Blog Details

Two words have quietly drained more value from AI projects than any technical failure ever will: good enough. It sounds prudent. It sounds like discipline. In practice, settling for good enough at launch is the most expensive AI decision you will make, because AI performance and optimisation is not a one-time bar you clear. It is a level you either defend or slowly lose.

Why is good enough so costly? Because AI performance is not static. The moment you stop improving a system, it does not hold steady at good enough. It drifts downward from there. So good enough at launch quietly becomes not quite good enough a few months later, and eventually not good enough at all, while everyone remembers signing off on something that was fine. You did not buy a stable outcome. You bought the highest point on a curve that only bends down without effort.

The comfortable lie inside good enough

Good enough feels like a mature engineering trade-off. Ship the thing that works, stop gold-plating, move on. For conventional software that logic holds, because the software does the same thing next year. For AI it quietly fails, because the ground moves. Good enough assumes a fixed target. AI lives against a moving one.

So the phrase smuggles in a false assumption. It treats launch quality as a floor you will stand on. It is actually a ceiling you will fall from. The decision to stop at good enough is really a decision to accept steady decline, whether or not anyone framed it that way.

How small gaps become large costs

Imagine a system that is good enough, meaning it handles most cases acceptably and fumbles a minority. That minority is where the cost lives, and it grows.

• The cases it fumbles are often the valuable or sensitive ones: the unusual customer, the high-stakes decision, the edge case that matters most.
• Each fumble carries a cost in rework, lost trust, or a worse outcome, and those costs accumulate quietly across thousands of interactions.
• As the system drifts, the share of fumbled cases grows, so a gap that was tolerable at launch widens into one that is not.

Good enough, in other words, is not a fixed level of acceptable imperfection. It is a starting point for a growing problem. The imperfection you accepted on day one is the smallest it will ever be.

The compounding advantage of not settling

Now flip it. A system that is actively optimised does not just avoid decline. It improves. Small, regular gains compound the same way small, regular losses do, only in your favour. The gap between a tended system and a neglected one widens every month, and after a year the difference is not marginal. One team is running well above launch quality. The other is running well below it. They started at the same place.

This is the real argument against good enough. It is not that launch quality is bad. It is that treating launch as the finish line forfeits all the improvement that disciplined AI performance and optimisation would have delivered, while also failing to defend the quality you started with. You lose twice, once on the downside you accept and once on the upside you never claim.

Where good enough is genuinely fine

To be fair, not everything deserves relentless optimisation. Some systems are low-stakes, low-volume, and stable, and for those good enough really is enough. The mistake is applying that logic by default to systems that are high-volume, high-stakes, or sitting in a fast-changing environment. The higher the volume and the stakes, the more a small quality gap costs, and the less good enough can be trusted to stay that way.

The discipline is not to chase perfection everywhere. It is to be honest about which systems can coast and which cannot, and to resist the comfortable habit of letting everything coast because coasting is cheaper this quarter.

The decision nobody remembers making

The most insidious thing about good enough is that it is rarely a conscious decision. Nobody holds a meeting to decide the AI should slowly get worse. It happens by omission. The project launches, attention moves elsewhere, no one owns ongoing performance, and the default takes over. The default is decline. Good enough is what that decline feels like on the way down, right up until it is not good enough and someone finally asks what happened.

Avoiding this does not take heroics. It takes deciding, on purpose, that the system has an owner, a cadence, and a small budget for staying sharp. That single decision is what separates the AI that keeps paying back from the AI that quietly stops.

What good actually costs

The objection is always cost. Surely relentless optimisation is expensive? In practice the opposite is true. Continuous tending is cheap because it works in small increments and catches problems early. It is neglect that is expensive, arriving as lost value and heavy recovery projects. Choosing good enough to save money usually spends more of it, just later and less visibly. Across more than 670 technology projects since 2022, the systems that stayed valuable were rarely the cheapest to launch. They were the ones somebody kept sharp.

How to escape the good enough trap

Escaping the trap is not about heroics or big budgets. It is about installing three small things at launch, before attention drifts elsewhere and the default takes over.

An owner. One named person accountable for whether the system still performs. Not a committee, not the whole team, one person who will notice and care. Systems without an owner do not get tended, because tending is nobody's job by default.

A cadence. A fixed rhythm for checking performance, whether monthly or quarterly, put in the calendar so it happens whether or not anything looks wrong. The whole point is to catch problems before they announce themselves, which means looking when nothing seems amiss.

A small standing budget. A modest, agreed amount set aside for tuning and refreshes, so the fix does not require a fresh business case every time. When keeping the system sharp needs a fight for funding, it loses the fight, and decline wins by default.

Those three things cost very little and change the trajectory entirely. They turn optimisation from an afterthought that never happens into a normal operating habit. The organisations that get lasting value from AI are rarely the ones with the most impressive launch. They are the ones who quietly built these habits in from the beginning and never let good enough become the plan. The difference does not show up on day one. It shows up on day three hundred, when their system is still sharp and everyone else's has quietly rotted.

The bottom line

Good enough is a trap disguised as prudence. It assumes a fixed target in a world that moves, treats a ceiling as a floor, and quietly commits you to decline you never chose. The alternative is not perfectionism. It is deciding that performance has an owner and a rhythm, so the system defends its quality and compounds small gains instead of shedding them. Settle at launch and you pay for it later. Keep it sharp and it keeps paying you back.

Talk to us before your AI starts to slip

If you are not sure whether your AI is still performing the way it did at launch, that uncertainty is itself the signal to check. Freemansland offers a free AI opportunity assessment where we give you an honest read on where AI helps, where it does not, and what it would take to keep a system sharp over time. No jargon, no obligation. Get in touch for your free assessment and we will come back within one working day.

Frequently Asked Questions

Is aiming for good enough ever the right call for AI?

Yes, for low-stakes, low-volume, stable systems where the cost of any imperfection is small and unlikely to grow. The problem is applying that mindset to high-volume or high-stakes systems in changing environments, where a small quality gap compounds into real losses. The skill is telling the two apart honestly rather than defaulting to good enough everywhere.

Why does AI quality decline if I do nothing?

Because AI makes decisions based on patterns in data, and both the data and the surrounding world keep changing after launch. As reality drifts away from what the system was built for, its accuracy falls. Without active monitoring and tuning to close that gap, the decline is gradual, continuous, and easy to miss until it becomes costly.

Does optimising AI really cost less than leaving it alone?

In most cases, yes. Continuous optimisation works in small, cheap increments and catches issues while they are minor. Neglect appears free but accumulates hidden losses and eventually forces expensive recovery work. When you compare the full cost over a year or more, tending the system is usually the cheaper path, not the more expensive one.