Malay Mehta
Back to Blog

By Malay Mehta, Software Architect and Mentor · 11 min read

AI Productivity Coach: What 60-70% Faster Actually Looks Like

The stories from my own work are real, with client and system names removed. The examples for HR, accounting and the other roles further down are made up to show the pattern. They are not client stories.

People keep asking me some version of the same question. Which AI tool should I use? Honestly, that's the wrong question. The tool matters less than you think. What matters is whether you know your job well enough to tell when the AI is wrong.

Here's my own number. In my day to day work as an architect, AI cuts the time on complex work by 60 to 70%. On simple and moderate tasks it's closer to 80%. That's real, I measure it against how long the same work used to take me, and I'm not going back.

But that number comes with a condition nobody puts in the marketing. It's 60-70% faster for me because I already know what good looks like in my work. Take that away and the number goes negative, because now you are fast at producing things that are wrong.

AI Multiplies Your Expertise. It Doesn't Replace It.

Think of AI as a multiplier. If you're a 9 out of 10 at your job, AI lets you produce like a small team. If you're a 2, AI helps you produce a lot of 2 quality work, very fast, with a lot of confidence. And it'll look polished, which is the dangerous part.

AI won't add new skills to your expertise. You still have to learn your craft, the hard way, same as before. What AI changes is how much of your day goes to typing, searching, formatting and first drafts. That time comes back to you. Whether you can spot the moment the output goes wrong is still on you.

That's the whole idea behind how I coach people on AI. I don't teach prompts. I teach where the human has to stand in the loop, for your specific job.

What 60-70% Faster Looks Like in My Week

Some concrete examples from the last few months.

  • Performance investigation. Two days of load testing a multi-tenant platform. The agent ran 12 load test cycles, pulled monitoring data, read CPU profiles and proposed causes. The test runs didn't get any faster. The thinking between them went from hours to minutes. Response time went from 14 seconds to 149ms. Full write-up here.
  • Infrastructure changes. Planning a Kubernetes cluster upgrade across two environments. AI wrote and checked the runbook. I ran it, step by step.
  • Testing an accounting product. AI wrote and ran hundreds of end to end tests, then compared a real customer's migrated data against their old system, document by document.
  • Planning and writing. Runbooks, Jira stories, design notes, cost breakdowns. The boring 80% of the document is done in minutes. The 20% that needs judgement is still mine.

The tools, since people always ask. Claude Code and Codex for work inside a codebase, where the agent reads, changes and runs things. Copilot for suggestions while I'm typing code. For a quick script or anything technical I go to Claude. For generic stuff like a quick analysis, a document draft or summarising something long, I often just use ChatGPT. MCP (Model Context Protocol) is what lets these tools talk to the systems I already use, like Jira, so the AI reads and updates tickets instead of me copy pasting. And agentic workflows tie it together: the AI does a multi step task end to end, and I approve the steps that matter.

None of those tools know when to stop, what's in scope, or what my client actually cares about. That part didn't get faster. It got more important.

Five Mistakes I See People Make With AI at Work

These come up in my own work all the time, and I'd bet most of them are happening in yours too.

1. Treating the output as a verdict

AI answers with the same confidence whether it's right or wrong. During the performance work, the agent read a database connection setting from the code, saw 5, and reported it as fact. Production had overridden it to 15. Its analysis was built on the wrong number. I caught it because I knew the environment. The AI only knew the code.

Another one. We were migrating a customer's accounting data from their old system. The AI read a "place of supply" field the same way on purchases as on sales. On purchases that field means something different. Result: 129 purchases with the wrong tax split. It got caught only because I insisted on comparing every migrated document against the old system's own totals before trusting any of it.

Treat AI output like a draft from a very fast junior colleague who never says "I'm not sure" unless you ask.

2. Not asking questions back

The most useful thing I do with AI is argue with it. While designing a data pipeline, the AI proposed building a separate component just to hand out sequence numbers. I asked, do we really need this, can't a timestamp do the job? Most of the time the simpler option is fine, and AI won't offer it on its own because it doesn't know what you are willing to trade off. You don't always need an axe where you need a knife. AI loves the axe.

Same in the performance work. The agent sized some new autoscalers from peak CPU instead of average, which made them 7 to 12 times too high. That was the exact bug it was supposed to be fixing. One question from me, "don't you think you needed to change inventory-service HPA?", made it go back and re-check everything. It ended up fixing five autoscalers instead of three.

3. Giving AI more control than you meant to

When I upgraded a production Kubernetes cluster with AI, my first message set the rules. Read-only commands in production. Ask me before every step. Record every failure, don't hide it under the carpet. Without that, an agent will happily "fix" things you never asked about.

It still tried. On a networking change, I asked for a fixed outbound IP address for a vendor. The plan came back also making the load balancer multi-zone. Sounds like an improvement, right? It wasn't in scope, we weren't running multi-zone, and that's a separate project with its own planning. An unasked change in production is a risk, not a bonus.

And in testing: a test failed because a list screen didn't refresh. The AI changed the test so it would pass, instead of fixing the app. Then it blamed browser caching, which was also wrong. I pushed back, "we own this code, it's our bug", and only then did it find the real problem. If you let AI decide what "done" means, it'll pick the easiest definition.

4. Not knowing the difference between plan and execution

A plan is cheap and you can throw it away. Execution isn't. I always ask for the plan first, review it properly, then let it run, often one phase at a time.

In the performance work the agent proposed two changes together. I said "make only pool changes to begin with", so we would know which one actually helped. At the end I decided when to stop: this is max capacity with these resources, and that's fine. AI has no sense of good enough. It'll keep optimising forever if you let it.

Smaller example. AI generated a set of Jira stories with acceptance criteria and no description at all. They looked complete. No developer could have picked one up and started. Read the plan as seriously as you read the result.

5. Starting a new project with no guardrails

Building something new from zero is where AI looks most impressive and does the most damage. No written spec, no clear boundaries between parts of the business (billing, inventory, users, each with its own rules about what it can touch), no tests from day one. So AI fills every gap with a guess. Each guess is reasonable. Together they're a mess by month three. In one service I work on, AI added code for something that belongs to a completely different system, because nothing told it where that boundary was.

In the accounting product, every feature had written requirements, tests and checks before AI touched it. Even then the test suite was green, and the first real company I opened in the app had problems within minutes. Guardrails reduce mistakes. They don't remove the need to look. I wrote more about this in why vibe coded apps break in production.

What This Looks Like in Your Job: HR, Accounting, Sales and More

You don't need to be technical for any of this to apply. The examples below are made up, but they follow the same patterns I see in my own work. For each one: what AI does well, where it goes wrong, and what the person who knows the job does about it.

HR and recruiting

AI screens 200 CVs in minutes and drafts job descriptions that read well. You ask it for the 20 strongest candidates. The shortlist quietly ranks down people with career gaps, parents coming back to work for example, because gaps look like risk. Nobody asked it to do that and nobody checked. A good recruiter sets the criteria first, asks the AI why each person was dropped, and spot checks the rejects. The hiring decision is yours, legally too.

Accounting

AI matches three months of bank transactions to invoices in an afternoon. A customer paid part of an invoice plus an advance for their next order in one transfer, and the AI marks the invoice as fully paid. The books look clean, but receivables and tax are wrong. An experienced accountant asks one question before accepting anything: show me every match where the amounts aren't exactly equal.

Learning and development

AI builds a six week leadership course with slides and quizzes in a day. Nobody checked what the managers actually lack. The course is polished, pitched at first time managers, and the audience is senior managers who sit through it bored. The fix is planning before production: who is this for, what can't they do today, what should change after. Then let AI build it.

Software development

AI builds a whole booking feature from one prompt over a weekend. Nobody wrote down what happens with cancellations, time zones, or two people booking the same slot. The demo works. In month three, double bookings. A senior developer writes those rules first, builds in small pieces, and has tests for the edge cases before the code exists.

Software testing

AI writes 300 tests overnight. The ones that fail get "fixed" by loosening what they check. Coverage looks great and the tests prove nothing. This is exactly what happened to me above. A good tester sets one rule: a failing test is information, and AI never changes a test to make it pass without explaining why the test was wrong.

Sales

AI personalises outreach emails for 50 prospects. One email congratulates a company on a funding round that never happened. The AI made it up to sound personal. A good salesperson checks every fact about a person or company against a real source before hitting send. One wrong fact and you've lost that prospect.

Project management

AI turns a kickoff meeting transcript into a project plan with a timeline. It puts design and build in parallel, when build can't start until design is signed off. The timeline looks four weeks shorter than reality and everyone commits to it. A good project manager asks AI to list every dependency explicitly and checks them against how the team really works.

Customer support

The AI is connected to the help desk through MCP and drafts replies. Someone gives it permission to also issue refunds, to save time. A polite but persistent customer gets a refund they weren't entitled to, and the same thing happens 40 more times before finance notices. The rule a good support lead sets: AI drafts, a human approves anything that moves money.

Marketing

AI writes a month of LinkedIn posts in an hour. Every post has the same rhythm and the same hooks, and readers notice. I learned this one myself. People on Reddit called some of my more polished writing "AI generated". Use AI for research and structure. Keep the words in your own voice, and edit until it sounds like you talking.

Notice that in every single one of these, the AI did 80% of the work well. The damage came from the 20% where someone who knows the job should have been looking, and wasn't.

How to Be the Human in the Loop: A Short Checklist

  • Know what good looks like before you ask. If you can't judge the answer, learn that part first.
  • Set the rules in your first message. What AI can do on its own, and what needs your approval.
  • Ask for the plan first. Review it. Then execute in small steps.
  • Ask "how do you know?" Make it show you where a number or fact came from.
  • Give it what it can't know on its own: your company, your customers, the exceptions nobody wrote down.
  • When something important is at stake, change one thing at a time.
  • Decide when it's done. AI doesn't know what good enough means for you.
  • Anything that touches money, people or production stays behind your approval.

Want Help Using AI in Your Own Work?

This is something I now offer alongside my engineering mentoring: one-on-one AI productivity coaching, for any job, not just software. Bring your real work. We pick one or two tasks you do every week, choose the right tool for them, and build a workflow where AI does the heavy lifting and you stay in control of the parts that matter. Same rate as my mentoring sessions, $60 an hour.

AI won't make you an expert. It will make an expert very fast. Put your time into both.

Want AI to save you real hours without losing control?

Book a session on Codementor, or email me at hello@malaymehta.com with a line about your job and what you want AI to help with, and we'll set up a time. 726+ sessions, 5.0 rating.

Get notified when I publish new posts

Architecture, system design, and production engineering.

Subscribe via Email