Delivery
How long does software take to build?
Why software estimates are wrong in a consistent direction, what actually determines a timeline, and how to plan around uncertainty instead of pretending it away.
Key takeaways
- Estimates are wrong in a consistent direction because the unknown work is always additional, never subtractive.
- Decision latency on your side is frequently the largest single delay, and it is the one nobody puts in the plan.
- Ask for ranges with the reasons attached. A single confident number is less information, not more.
- Track how long things actually took against what was estimated. Two months of that data beats any estimation technique.
- The reliable way to go faster is to build less. Everything else is a marginal improvement on the same work.
Every founder has been told six weeks and received four months. It is tempting to conclude that engineers are bad at estimating, or optimistic, or being deliberately vague.
The real explanation is structural, and once you see it the planning problem becomes tractable.
Why estimates skew one way
If estimates were merely imprecise, they would be wrong in both directions and average out. They do not. They are consistently short, and the reason is that the work you discover is nearly always additional.
You find out the API you are integrating with has no sandbox. You find out two requirements conflict. You find out the client's data has a shape nobody mentioned. You almost never discover that something is unnecessary.
So the error distribution is one-sided. Any estimate produced without accounting for that is a best case being reported as an expectation.
What actually determines the timeline
Scope, obviously. But specifically the scope nobody wrote down. Undocumented assumptions become work at the point somebody notices them.
Integrations you do not control. Access, credentials, sandbox environments, undocumented behaviour, support tickets that take a fortnight. This is the most under-estimated category in almost every plan.
Decision latency on your side. This is the one nobody puts in the plan and it is frequently the largest. If a question takes five days to answer because the person who can answer it is busy, and there are twenty such questions, that is a hundred days of partial blockage in a project you estimated at sixty.
How well the ground is prepared. Environments, access, test data, a person who can answer questions. A team waiting for a VPN account is a team not building anything.
Requirement conflicts. Two stakeholders wanted incompatible things and nobody noticed until it was implemented.
Notice that "how fast engineers write code" is not on this list. It is the part everyone optimises and it is rarely the constraint, which is also why adding people to a late project reliably makes it later.
What AI changed, and what it did not
Implementation is faster. That is real, and it has genuinely compressed the build portion of a timeline.
Be careful about extrapolating it to the whole. METR's randomised trial found experienced developers were 19% slower on tasks in codebases they knew well while believing they were 20% faster, a useful caution against planning from perceived speedups. DORA's 2025 data found AI adoption correlating with higher throughput and higher instability together: faster delivery, more rework.
The practical planning consequence: the code-writing portion has shrunk, everything around it has not, and some of the gain returns as review and rework. Expect a meaningfully shorter build phase, not a shorter project by the same proportion.
Rough shapes
Assuming a competent team and decisions that arrive within a day or two:
| Work | Typical |
|---|---|
| Scoping and specification | 1–2 weeks |
| Minimal product, one workflow | 4–8 weeks |
| Multi-role product | 3–6 months |
| Integration with a legacy system | Add 2–6 weeks per system, high variance |
| Compliance and security review | 2–6 weeks, often in parallel |
| Launch hardening | 1–2 weeks, always needed, usually forgotten |
That last row is worth noting. The gap between "it works" and "it is ready for real users" is consistently underestimated, because it consists of unglamorous work, error states, empty states, monitoring, the sign-up flow, that nobody demos.
How to plan honestly
Ask for a range with reasons. "Six to ten weeks, and here is what determines which end" contains far more information than "eight weeks".
Ask what would make it wrong. This surfaces specific risks you can act on, rather than a buffer you cannot inspect.
Track actuals against estimates. Two months of this data tells you your team's specific bias, most teams are consistently off by a stable factor, and knowing yours is worth more than any estimation methodology.
Re-estimate at milestones. An estimate made when 20% is complete is dramatically better than one made at zero. Plan to revise rather than treating revision as failure.
Measure your own decision latency. How long between a question being asked and answered? Most clients are surprised, and it is the delay most within your control.
The only reliable lever
Build less.
Everything else on this page is a marginal improvement to how accurately you predict the same amount of work. Cutting a user type, deferring the admin panel or doing onboarding by hand for three months changes the amount of work, and it is available immediately, at no cost, on every project.
Frequently asked questions
Why are software estimates always wrong?
Because estimating requires knowing what the work is, and a large share of the work is discovered while doing it, the integration that is undocumented, the edge case nobody mentioned, the requirement that turns out to conflict with another. That discovery is asymmetric: unknown work is nearly always additional. This is why estimates are wrong in one direction rather than scattered evenly around the truth.
How long should a typical MVP take?
Four to eight weeks from an agreed specification to a deployed product with real users, for a genuinely minimal scope. The variable is rarely engineering speed; it is how quickly scope can be cut to one thing that must work well, and how quickly decisions get made once questions arise.
What actually causes software projects to run late?
In rough order: scope that grew without anyone deciding it should, integrations with systems nobody controls, decisions waiting on someone's availability, and requirements that turn out to conflict. Notably, engineers writing code more slowly than expected is well down the list, which is why hiring more of them rarely fixes a late project.
Should I add a buffer to a supplier's estimate?
Rather than a flat multiplier, ask what the estimate assumes and what would make it wrong. That surfaces the specific risks, which you can then either mitigate or plan around. A blanket buffer hides the risks instead of managing them, and it tends to get consumed regardless because work expands to fill it.
How can we make software get built faster?
Reduce scope, and decide faster. Those two account for most of the achievable difference. Beyond them: make sure environments and access are ready before work starts, keep the team small and uninterrupted, and ensure someone with authority can answer questions within a day rather than a week.
References & further reading
- [1]
- [2]
- [3]Martin Fowler, on estimation and its purpose
martinfowler.com
Related resources
How much does it cost to build an MVP?
An honest breakdown of what a first product version costs in 2026, what actually drives the number up or down, and how to tell a realistic quote from an optimistic one.
Fixed price vs time and materials: which contract to use
Both pricing models fail in predictable ways. How to tell which one fits the work you actually have, and how to structure either so incentives point in the same direction.
How I scope a build so it does not run over
The specification format I use before writing any code, what goes in it, what deliberately stays out, and why scoping is now the constraint rather than implementation.
Turn the trend into shipped software
I contract as a senior fullstack and data engineer, building the production systems your customers and your reporting depend on.