Every team that adds a machine learning feature to a product eventually faces the same awkward meeting. Someone wants it live next month, someone else has a fixed budget, and a third person expects answers that are right nearly every time. This explainer unpacks the trade-off between speed, cost and quality in AI projects, in plain terms, so that the people involved can agree early on what gets loosened, long before any engineering starts.
The rule of thumb behind AI delivery choices
The idea is old and simple: fast, cheap and good cannot all be maximised together. In an AI setting the three words take on specific meanings.
- Fast can mean quick to build, or quick to respond when a user asks a question. Those are different kinds of fast, and projects often confuse them.
- Cheap covers the compute bill, the data preparation hours and the people who maintain the system afterwards.
- Good means the output is accurate enough, consistent enough and safe enough for its purpose.
Pick any two and the third usually bends. A system that is quick to ship and inexpensive tends to be rougher around the edges. A system that is accurate and cheap tends to take longer, because someone has to clean data and tune carefully. A system that is accurate and quick tends to need more compute and more skilled hands, which costs money.
Where each corner shows up in practice
Giving up quality on purpose
Sometimes a rough answer is plenty. A tool that suggests tags for internal documents can be wrong now and then without harm, because a person glances at the tags anyway. In that case a small, inexpensive model delivered quickly is a sound decision. The key is that the team states, in writing, that mistakes are expected and tolerated, and decides in advance what a human will check.
This is quite different from a system that influences money or safety. There, relaxing quality is rarely acceptable, so the team has to give up one of the other corners instead. The business angle gets covered in a piece about the role AI plays in day-to-day company choices, and the lesson carries over: whatever rides on an answer should determine how much error is tolerable.
Giving up speed on purpose
Where accuracy matters and money is limited, the honest answer is to go slowly. That means more time collecting representative examples, more time reviewing mistakes, and a gradual rollout. The cost stays modest because you are spending calendar time rather than expensive hardware. The price is that stakeholders wait, and waiting is uncomfortable when a competitor seems to be moving faster.
Speed also has two senses. A model that answers in a fraction of a second at runtime is a different requirement from a team that ships in two weeks. You may want the first and not need the second. Separating them avoids over-spending on a requirement nobody actually has.
Giving up low cost on purpose
If a feature must be both quick to market and reliably accurate, money buys the missing piece: larger models, more labelled data, a bigger engineering team, extra testing. Accepting a higher bill is a legitimate decision as long as it is deliberate. The failure mode is the team that promises all three, discovers mid-project that the budget cannot support it, and then quietly cuts testing to compensate.
A worked example: a support-ticket classifier
Suppose a company wants software that sorts incoming support tickets into categories so that the right team sees each one. Here is how the three corners might play out.
- Quick and inexpensive. Use a lightweight model trained on a few hundred past tickets. It is live within days and runs on modest hardware. Expect it to confuse similar categories, so route low-confidence tickets to a person.
- Inexpensive and accurate. Spend several weeks on labelling and on checking edge cases, then deploy the same lightweight model. Accuracy improves because the training data is better, not because the model is fancier. Launch is later.
- Quick and accurate. Pay for a larger model and a short burst of contractor labelling time. It ships fast and performs well, and the invoice is larger.
All three are defensible. What is not defensible is choosing the first and telling management it will behave like the third.
Why it helps to say it out loud
Most disappointment in AI work comes from silent assumptions. A product owner imagines near-perfect answers; an engineer assumes a prototype is wanted. By stating “we are choosing speed and low cost, so expect imperfect output and a human review step”, everyone shares the same picture. Later arguments become about whether the choice was right, which is a healthy argument, instead of about who promised what.
It also changes how success is measured. If quality was the sacrificed corner, success is not perfection. It is whether the rough output saved time compared with doing the task by hand.
Reading about operations from another angle
Operations teams wrestle with this triangle daily, and other people’s descriptions of it can sharpen your own. One place that does so in short, plain language is a blog on cloud operations called CloudOps Now, which presents itself as a CloudOps blog and puts out brief articles on subjects like running costs for networks. You do not need to agree with every post to find the framing of cost against delivery useful when you plan an AI rollout.
A short checklist before you commit
- Write down what a wrong answer costs, in time, money and trust.
- Decide which corner is flexible: the date, the budget or the accuracy bar.
- Separate build speed from response speed.
- Name the safety net, such as a human check for uncertain cases.
- Tell stakeholders the choice before work begins.
When visual and data-heavy systems raise the stakes
The same logic applies to heavier workloads. A related article about scaling image-heavy systems while avoiding slow spots illustrates how pressure on responsiveness drives spending up, which is the triangle at work. Image and video pipelines often demand real-time responses, so the quick corner is fixed, and the team must choose between a larger budget and a more modest output quality.
What the sacrificed corner looks like in a roadmap
Once a choice is made, it should show up in the plan rather than live only in a meeting note. If quality is the corner being relaxed, the roadmap should include a review loop: a weekly sample of outputs checked by a person, a clear threshold at which the feature gets paused, and a date to revisit the decision. If speed is relaxed, the roadmap should show intermediate milestones, so that waiting stakeholders can see progress in the form of better data and clearer error analysis rather than a silent gap. If cost is relaxed, the plan should say what the extra spend buys and when it will be reviewed.
A plan that records the trade-off is also easier to hand to a new colleague. A newcomer learns what exists and also the reason a given corner was loosened, which prevents well-meaning attempts to fix something that was left rough on purpose.
Common mistakes to avoid
A few errors recur. Teams declare all three corners non-negotiable and then let one erode silently. Teams treat a prototype demo as proof that production quality is near. Teams forget that cheap to build is not the same as cheap to run, since a model that is inexpensive to train may be costly to serve at volume. Naming the trade-off early protects against each of these.
Summing up
AI projects are not exempt from the old project triangle. They simply express it in new vocabulary: latency, labelled data, compute, error rates. Choose the corner you are willing to relax, say so plainly, and design a safety net around it. A project that states its trade-off honestly is far more likely to land well than one that promises everything and quietly delivers two out of three.