Imagine a lunch-hour kitchen in Gordon’s Bay. The extractor hums, plates knock together, and someone has found a wonderfully fast way to prepare sandwiches. Bread, filling, wrap. Done. Another one. Done.
At the collection counter, though, nothing has changed. One person is still checking orders, sorting allergies and matching each bag to the right customer. The sandwiches pile up. People wait.
Would we call that kitchen more productive?
One part of it certainly is. Whether lunch reaches people sooner is another question.
That’s the distinction I want us to make when we talk about AI productivity. Not because I think the technology is useless. Quite the opposite. A genuinely useful tool deserves a better test than how impressive it looks while everyone downstream quietly drowns.
Faster at what, exactly?
There are three things we keep bundling together: activity, task speed and useful work delivered.
They aren’t interchangeable.
More prompts, more tokens consumed, more lines of code, more drafts: those tell us something happened. They don’t tell us whether it helped. Finishing an individual task sooner tells us more. But even that doesn’t settle whether the customer received a working result sooner, at a sensible total cost.
An invoice reminder drafted instantly but waiting days for permission to send is still waiting. A feature written quickly but built around the wrong requirement is quickly wrong.
These are illustrations, not stories about a particular client. But the distinction is practical enough for any small business here in South Africa. The customer doesn’t buy our typing speed. They buy something that works.
I like speed. I’m a technology entrepreneur; I’m hardly campaigning for everyone to go back to copying things by hand. What bothers me is the temptation to stop measuring at the exact point where the numbers become flattering.
Apparently the race ends where my part ends. Very convenient for me.
The research is more interesting than the headline
Dylan John’s video, “Harvard Just Discovered AI Isn’t Actually Working”, raises the useful question of what happens after code is generated.[4] But I don’t think that headline is a fair summary of the underlying research.
Fiona Chen and James Stratton’s working paper, Artificial Intelligence in the Firm: Bottlenecks in Software Production, studies 445 million work events covering 720,542 workers at 655 firms, from January 2021 to August 2026.[1] I’m referring to its 14 November 2026 version.[1]
It isn’t a randomised trial, or an institutional verdict that AI has failed.[1] The researchers compare firms adopting at different times.[1] Their causal interpretation depends on an important assumption: without adoption, earlier and later adopters would have followed similar paths.[1]
The distinction they find is between intermediate code output and completed software work.[1] Twelve months after agent adoption, they estimate a 35.7% increase in pull requests opened per worker: proposed bundles of code submitted for review.[1] That agent estimate is additional to earlier assistant adoption, not a clean comparison with never using AI.[1]
For final output, their per-worker estimates are increases of 13.7% in resolved Jira issues and 18.0% in resolved epics, the larger projects or features.[1] Neither estimate provides firm statistical evidence of an increase; the issue result is borderline.[1] The earlier assistants’ final-output estimates were near zero.[1]
So, no: the current paper does not say agents produced literally zero extra completed work.[1] Nor should we treat those estimated gains as settled promises.
The gap is the point. Much more activity entered the process than emerged as additional completed software.[1] And even completed ticket counts are imperfect stand-ins for customer value; they aren’t a direct measure of whether anyone’s life or business improved.
The paper also estimates a 56.1% increase in time from code submission to approval after agent adoption.[1] Code waited longer before review and spent longer in active review.[1] Review comments increased too.[1]
That sounds less like “nothing works” and more like “we sped up one stage and discovered what was holding the rest together”.
There are limits. These were largely established, medium-to-large firms, not a representative sample of every small team.[1] I wouldn’t paste their percentages onto my own business and call it a forecast.
I would take the question seriously, though.
Where is the work waiting now?
The queue didn’t disappear. It moved.
Back to our imaginary kitchen. Buying a faster sandwich press doesn’t increase the person at the counter’s capacity to check orders. If we keep feeding that counter faster, we may simply create a larger pile.
Software has its own version. Someone must understand the requirement, review the change, test the awkward cases and decide whether it’s ready. The paper finds review becoming a constraint; it doesn’t establish that every surrounding task got slower, or that every AI-generated change is worse.[1]
That matters. More comments aren’t automatically proof of rubbish code. They are evidence of additional review activity. The researchers interpret the patterns as consistent with both capacity pressure and greater review effort, rather than directly inspecting every line for quality.[1]
Outside software, I’d treat this as a question to investigate, not a finding already proved for every profession. Does the faster draft leave more checking? Does a quicker proposal create a longer approval queue? Does the person who knows the customer now spend the day explaining what the generated work misunderstood?
And sometimes the bottleneck is embarrassingly close to home: the founder insisting on approving everything.
If every ordinary decision needs me, perhaps I shouldn’t be quite so pleased that everyone can now produce more decisions for me to make.
A faster tool can expose a management problem. It cannot accept responsibility for fixing it.
Make the work easier to finish
My first response wouldn’t be to demand faster approvals. That can turn a useful safety check into a tired person clicking yes.
I’d start before generation. What needs to change? For whom? What would demonstrate that it works? Who is allowed to accept it?
Imagine a checkout repair. “Improve checkout” leaves a lot to argue about afterwards. “Let a customer complete this purchase on a phone without losing their delivery details” gives the implementer and reviewer something concrete to examine.
Then keep the change small enough to understand. One repair, its relevant tests and a plain explanation are easier to judge than the same repair bundled with a redesign nobody requested. This is my practical recommendation, not an experiment the paper ran.
Use AI to help produce that explanation or explore test cases, certainly. But a confident summary isn’t independent evidence that the change works. The reviewer still needs to inspect the relevant behaviour and challenge the assumptions.
The checks should match the risk. A heading correction shouldn’t face the same process as changing payment handling or access to customer information. Equally, “we’re moving fast” isn’t a reason to wave the dangerous change through.
I’d also protect time for review, make ordinary decision rights clear and reduce the number of unfinished jobs competing for attention. Starting another thing can feel productive when finishing the current thing needs an uncomfortable conversation.
Sometimes the useful next action is not another prompt. It’s answering the question someone asked yesterday.
My AI productivity measure: the Finish-Line Test
Here’s the question I’d put in front of myself before celebrating a gain:
What useful outcome was accepted and delivered, with verification, rework and the real end-to-end cost accounted for?
That’s the Finish-Line Test. Not another leaderboard.
Take one recurring piece of work and define its finish before trying to accelerate it. For that imaginary checkout repair, “code generated” isn’t the finish. A verified purchase journey, accepted by the responsible person and available to customers, is much closer.
Measure the whole journey: when the clear request became ready, how long the work waited, how long implementation and checking took, and what had to be corrected afterwards. Include the person answering questions, the reviewer’s time, tool charges and any recovery work. Otherwise I can save my hour by quietly spending yours.
Then compare reasonably similar work before and after the change. A tiny wording fix and a complicated payment repair aren’t a fair comparison just because both count as one task. Keep the quality and acceptance standard consistent too. Finishing faster by dropping the check is a different experiment.
Don’t make employees build an elaborate reporting system to prove that reporting got easier. A modest record of completed outcomes, waiting time, rework and total effort can start a useful conversation.
If delivery improves without hiding defects or shifting an unreasonable burden onto someone else, good. Keep what helped. If generation accelerates but accepted delivery barely moves, look at the queue before purchasing more generating capacity.
And if a better process returns time to people without increasing output, that can still be worthwhile. Less strain is not a fake benefit. Just name it honestly rather than claiming extra customer value you haven’t measured.
Accountability doesn’t move with the typing
For me, this isn’t only an operational question. It is an amanah question: what have I been trusted with, and what am I claiming to have done?
Allah says, in Dr Mustafa Khattab’s translation:
“Do not follow what you have no ˹sure˺ knowledge of. Indeed, all will be called to account for ˹their˺ hearing, sight, and intellect.” — Qur’an 17:36.[2]
That is a wider instruction about knowledge and accountability, not a verse about software metrics. My business application is this: I shouldn’t confidently announce a saving I haven’t checked, or repeat an alarming headline without respecting the evidence behind it.
That reminder includes this article. It includes me.
Allah also commands returning trusts to their rightful owners and judging fairly, in Qur’an 4:58.[3] I understand that as a reason to take the customer’s money, the employee’s effort and the responsibility of approval seriously. It is my application of the principle, not a newly invented ruling about every automation decision.
The Prophet ﷺ said:
“Surely, everyone of you is a guardian and responsible for his charges.” — Sahih al-Bukhari 7138, closing sentence.[5]
The narration concerns responsibility across different positions of care and authority. I take its reminder into leadership: delegating the typing does not mean delegating away my answerability.
If I require someone to approve work, I owe that process more than their name at the bottom. They need relevant competence, enough time and the genuine authority to say, “This isn’t ready.” Otherwise I have assigned blame, not enabled judgment.
That distinction feels particularly important when jobs enter the conversation.
The paper did not detect a statistically significant employment effect in its observed sample.[1] But its estimates are noisy, and its separate long-run model scenarios can predict employment declines.[1] Those simulations depend on assumptions about future progress, demand and review capacity; they aren’t observed layoffs, or a guarantee about anyone’s job.[1]
So I can’t honestly use the research to promise that nobody will be displaced. Nor can I use a faster coding demonstration as sufficient evidence that a particular employee is now unnecessary.
Businesses face real constraints. This isn’t a claim that every staffing change is wrong. It is a plea for evidence, fairness and basic human dignity when decisions carry consequences for other people.
And for enough humility to notice when the overwhelmed reviewer is not resisting progress. They may be the person preventing a confident mistake from becoming the customer’s problem.
What actually reached the customer?
I still want the tools. I want the tedious work reduced, the useful ideas explored and the good changes delivered sooner.
But I don’t want to build a business that produces work faster than it can understand what it is producing.
Maybe the next improvement is better automation. Maybe it is a smaller change, a clearer requirement or permission for someone competent to decide without waiting for me.
The pile at the counter won’t tell me which. I have to look at the whole journey.
May Allah make us honest about what we know, fair to those whose work we depend on, and careful with what we approve. And may He help me ask the less flattering question before I celebrate the faster machine:
What actually reached the person we were supposed to help?


