Skip to main content

TECH VEDA

Corporate on-site training - Submit proposal Pick your modulesSharpen your kernel skills: deep dives, drivers, Yocto, CVEs, careers — updated daily. Read the blog →
Career

Owning a Project: The Challenges AI Assistants Do Not Solve

Owning a Project: The Challenges AI Assistants Do Not Solve

Embedded systems consultants and engineers often own the complete software delivery for a customer’s board, and AI coding assistants now do much of the coding. Four responsibilities remain that the assistants do not cover: agreeing the task with the customer, reviewing the work, committing to a date, and reporting progress. This article explains how to handle each one, with example messages you can adapt.

You give your AI coding assistant a task: “Add retry handling to the I2C read function.” After a few minutes it replies: “Done. All changes are complete.”

You do not accept this reply. You ask which files it changed, whether it ran the tests, and what it did not check. You ask before you use the code, because “Done” does not tell you enough to decide anything.

Now think about your customer. On Monday the customer’s engineer wrote to you: “Please get the touchscreen working on the revision B boards. We need it for our demo.” You replied, “Yes, we will do it.” You are the only software engineer on this project. Today is Thursday. The driver still does not detect the controller at boot, and you have not written to the customer since Monday. This example is invented.

This is what owning a project means: the customer is now in the position that you were in with the assistant. The customer has one sentence from you and has to make plans with it.

Why one engineer now owns a whole project

In embedded work, one engineer often owns the complete software for a customer’s board: the bring-up, the drivers and the delivery. Many consultants work this way on every project. AI assistants make it more common, because one engineer can now produce code that needed two or three. This is my observation from embedded projects. I have no data on how common it is.

You wait for the assistant. The customer waits for you. What you need from the assistant, the customer needs from you.

The assistants now do much of the coding that the team did. They do not solve four other challenges. In a larger team, other people handled these for you, and often it was nobody’s official job. So nobody tells you that they are now yours.

Challenge 1: turning the customer’s request into a clear task

In a larger team, a lead spoke to the customer and gave you a clear task. Now the customer’s two lines come directly to you.

You do this already for your assistant. A good prompt states three things: the goal, the limits, and the result that will show the task is finished. Write the same three things for the customer’s request, and send them to the customer before you start.

“To confirm: single-touch input working on your revision B boards with the current software image, delivered by Thursday evening, so that you can test it before your demo on Friday. The task is finished when your demo application responds to touch on two revision B boards. Multi-touch is not included. Please reply if any of this is wrong.”

In the example, the customer’s hardware engineer might have replied: “Correct. Please note that the reset line changed on revision B.” The work of stating a task clearly before you type a prompt is explained in the article on problem framing.

Your move → For the next request from your customer, write the goal, the limits, and the result that shows the task is finished. Send these lines to the customer and wait for a written “yes” before you write any code.

Challenge 2: reviewing the work before the customer receives it

In a larger team, a senior engineer read your change before it was delivered. Now the first person who reads your work after you may be the customer.

With an assistant, code that appears to work is ready much sooner. Much of it was not written by you, and no second person has read it. In embedded work there is one more limit: the assistant cannot see your board. It cannot measure the reset line or watch a cold boot. It works only from the text that you give it. Why code that compiles and runs is not yet proof that it is correct is explained in AI Can Write Kernel Code. Only a Human Can Sign It Off.

When an assistant finishes, you want to know what it did, what it checked, and what it did not check. Write the same three things about your own delivery, and write them before you send it.

“Delivered: touch driver update for revision B. Changed: reset pin in the device tree, and the reset delay in the probe function. Tested: ten cold boots on two revision B boards, touch detected every time. Not tested: suspend and resume, and revision A with this image.”

If the “Not tested” line is long, test more, or agree with the customer what will stay untested. An assistant often does not state what it left unchecked. You can state it, and I explained why this is valuable in Your Code Got Cheap. Your Judgement Did Not.

For the part with the highest risk, ask a colleague from another project to read it.

Challenge 3: deciding what date to promise

In a larger team, the customer asked the manager for a date, and the manager answered after asking you. Now the customer asks you during a call: “Will Friday work?” A date that you say during a call is a promise, and no manager will correct it afterwards. So do not answer with a date during the call. Say what you know, and say when you will send the date.

“We are still testing touch input on revision B. I will check two things after this call, and I will send you the date in writing by 5 pm today.”

Then send the date in writing, with the condition that it depends on: “Thursday evening, if the cause is the reset pin change that your hardware team confirmed.”

A technical date is different from the contract. If the customer asks for work outside the agreed scope, or for a change in price or in a contract date, do not agree during the call. If you are an employee, tell your manager the same day. If you are a consultant, agree the change in writing before you start the work. Raise any concern about safety, security or compliance in the same way.

Challenge 4: reporting when nobody asks

In a larger team, somebody asked for your status every day. Now nobody asks.

When an assistant runs for a long time with no output, you do not know whether it is making progress. The customer has the same doubt about you, but usually does not ask. When you send no update, the customer assumes that there is no problem and continues to plan for Friday.

So send a short written update on a fixed day every week, and send one more on any day when the date or the risk changes.

“Touch is not working on revision B yet. The controller is not detected at boot. I am checking the reset and interrupt pins against your revision B schematic today. Did any pin assignments change between A and B? Next update: 12 noon tomorrow.”

What happened when people trusted the whole workflow to an AI tool

The four challenges become dangerous when you give them to the assistant too. The four public cases below are single cases, not statistics, and the four challenges themselves are my own argument from practice.

An agent made a wrong statement about its own work. In July 2025, a user of the coding platform Replit reported that its AI agent deleted his database during a period when he had told it to make no changes. The agent then told him that the deletion could not be reversed. This was wrong, and the data was later restored. The company’s chief executive said publicly that the incident was unacceptable and that backups existed. An agent’s statement about what it did, and about what can be recovered, is not a verification.

An outside reader found the errors after delivery. In 2025, Deloitte delivered a report to an Australian government department under a contract of 440,000 Australian dollars. After publication, a university academic found references to documents that did not exist. The corrected report stated that a generative AI tool had been used, and the firm agreed to repay the final payment of the contract.

A company was held responsible for what its tool told a customer. In February 2024, a tribunal in British Columbia, Canada, decided a claim against Air Canada. The airline’s website chatbot had given a customer wrong information about a refund. The airline argued that it was not liable for what the chatbot said. The tribunal rejected this, said that the airline was responsible for all the information on its website, and ordered it to pay the customer 812.02 Canadian dollars.

People sent reports that they did not understand. In January 2026, the curl project stopped paying rewards for bug reports. It still accepts security reports. Its maintainer wrote that the number of poorly researched reports, generated by AI or not, was too high a load for the security team. His rule was that you should never report a bug unless you understand it and can reproduce it.

The customer holds you responsible for the delivery, not the tool.

What the assistant may prepare, and what must come from you

  • The task. The assistant may draft the confirmation message. You must read the customer’s request yourself and decide the limits.
  • The review. The assistant may write the code and the tests. The “Tested” and “Not tested” lines must come from tests that you ran or watched. A review by the same assistant that wrote the change is not a review. Also, do not give an agent permission to delete or change anything that you cannot restore.
  • The date. The assistant may list the remaining work. The date is your promise, so you decide it.
  • The report. The assistant may correct your language. Every fact in the report must be one that you checked.

If the customer asks a question about your delivery and you cannot answer it without the assistant, you do not yet understand what you delivered. I wrote about this problem in The Code Is Yours. The Understanding Is Not.

Where the comparison with an assistant is not true

An assistant is not afraid of blame. You are. When you work alone, it is easy to believe that a problem can stay hidden. But the customer will notice. A delay that you report on Tuesday is a change of plan. The same delay found by the customer on Friday is a failure. Why staying quiet seems safer is the subject of Why Growth Mindset Advice Works for Some Engineers and Not Others.

An assistant accepts every instruction. You should not. Sometimes the customer will ask for something that you think is wrong, for example to deliver without the suspend and resume test. State your objection one time, with evidence, in writing. If the customer still wants it, do the work properly, and record in the delivery note what was decided and by whom.

State your disagreement one time, with evidence. Then do what was decided, and keep a written record.

How to check that it is working

Check these every week.

  1. You have a written “yes” from the customer for the goal, the limits and the finished result of your current task.
  2. Your last delivery note had a “Tested” line and a “Not tested” line.
  3. You gave no date during a call. Every date went to the customer in writing, with its condition.
  4. The customer did not have to ask you for a status this week.
  5. Every change of scope, price or contract date was agreed in writing, or reached your manager on the same day.
Your move → Write the four challenges on one page: clear task, review, date, report. Next to each, write the name of the person who handled it in your last team. Then write what you will do this week in place of that person.
Was this worth your time?

Frequently asked questions

How often should I update a customer when I work alone on the project?

Send one short written update on a fixed day every week. Send one more on any day when a date or a risk changes.

Should I tell the customer that I used an AI assistant to write the code?

Follow the contract, and your company’s policy if you are an employee. In every case, tell the customer what you tested and what you did not test. That information helps the customer more than knowing who typed the code.

Who reviews my code if I am the only engineer on the project?

You review it first, by writing what you tested and what you did not test. For the part with the highest risk, ask a colleague from another project to read it. A review by the assistant that wrote the code is not enough.

RB
Raghu Bharadwaj

Founder, TECH VEDA — 20+ years teaching the Linux kernel, device drivers and embedded systems.

Follow on LinkedIn

Get new posts by email

Kernel, embedded Linux and AI-era engineering — a few sharp reads a month. No spam.

We email occasionally and never share your address.