Software Project Estimation: How to Present a Range Finance Can Plan Against
An engineering leader asked how long it will take to replace a regional bank’s legacy reporting system usually gives one number, and finance writes it down as a commitment. Nobody yet knows how many of the system’s reports are still used, how much business logic is buried in stored procedures or whether the new numbers will tie out to the old ones. Software project estimation is the practice of forecasting effort, duration and cost together with an honest statement of confidence, then narrowing that forecast as the unknowns get resolved. The estimate a CFO can plan against is a range, the assumptions behind it and the date it will narrow.
What Is a Software Estimate, and What Is It Not?
A software estimate is a forecast of how much work something will take, given what the team knows today. It is different from a target, which is what the business wants, and from a commitment, which is what the team agrees to deliver. Mixing the three is where most budget arguments start: the CFO hears a commitment, the engineers meant a forecast and the board approved a target.
An estimate is not a quote, a plan or a promise. It carries uncertainty by definition, and an estimate presented without that uncertainty is still uncertain; it has only hidden the part finance most needs to see. Separating the three in the first conversation (“our forecast is this range, the target you’ve set sits here inside it, and here is what we can commit to now”) gives everyone the same picture.
Why Are Early Estimates So Wide?
Early estimates are wide because the decisions that determine the work have not been made yet. The cone of uncertainty describes this: the range of likely outcomes is widest at the start of a project and narrows as requirements, design and technical choices get settled. The idea traces to Barry Boehm’s work on software cost estimation, and Steve McConnell popularized it under that name. The cone narrows only when decisions are actually made. A project that keeps its requirements open stays wide no matter how much time passes.
Consider adding single sign-on to a B2B SaaS product. At the first conversation, the team does not know whether customers need SAML, OpenID Connect or both, whether they also want automated user provisioning, or how existing users who signed up with passwords will be linked to their company identity. Each answer can change the work substantially. Once the product manager confirms which identity providers the first three enterprise customers use and whether provisioning is required, the range can narrow sharply, without anyone working faster.
How Do You Decompose Unknowns Before Estimating?
Break the work into three kinds: work the team understands, questions the team knows it has to answer, and risks nobody can name yet. The first kind can be estimated directly. The second kind needs investigation before it can be estimated with any confidence. The third kind is covered by contingency and by checkpoints that will surface it early.
For the bank’s reporting system, the known questions are concrete. How many reports exist, and how many ran in the last few quarters? Which ones feed regulatory filings such as the quarterly Call Report, and so need exact reconciliation? How much logic lives in stored procedures and spreadsheets rather than the reporting tool? Query logs and a conversation with the finance team can answer the first question in days, and the answer often shrinks the scope, because many legacy reports turn out to be unused.
The SSO project has its own list: the protocols each customer’s identity provider supports, whether provisioning and deprovisioning must happen automatically through SCIM or just in time at login, how an organization enforces SSO for all its users, how an administrator recovers access after a misconfiguration and how signing certificates are rotated. Each item is small to state and can be large to build, which is why each deserves its own line in the estimate.
Dependencies outside the team belong on the list too, because they move the calendar even when they add few hours. A bank bringing in an outside engineering partner will usually run a third-party risk review first; U.S. banking agencies issued joint guidance in 2023 on managing risk across the life cycle of third-party relationships. Source: Federal Reserve, SR 23-4: Interagency Guidance on Third-Party Relationships An enterprise SaaS customer may run its own security review before enabling SSO in production.
How Do Spikes Reduce the Biggest Unknowns?
A spike is a short, time-boxed investigation that answers one estimation question with working evidence instead of opinion. Its output is an answer and a revised estimate, not production code, and it is worth running whenever one unknown dominates the range.
For the bank, a useful spike rebuilds a small sample of representative reports end to end, including one that feeds a regulatory filing, and reconciles the totals against the legacy system. If the sample ties out cleanly, the per-report effort becomes something the team can multiply. If it does not, the team has found undocumented logic early, while it can still be estimated, rather than during a parallel run near cutover.
For the SSO project, the spike connects a test tenant to the identity providers the first customers actually use and walks through login, a new user’s first sign-in and the removal of a user who has left the customer’s company. Deprovisioning is often the step that turns a simple login feature into a larger piece of work, because a person who has left a customer’s company must lose access promptly.
| Unknown | Spike Question | What the Answer Changes |
|---|---|---|
| Report inventory | Which reports ran recently, and who uses them? | Total scope, often downward |
| Embedded business logic | Do sample reports reconcile to legacy totals? | Effort per report and the size of the parallel run |
| Identity provider mix | Which protocols do the first customers’ providers use? | Whether one or two protocol implementations are needed |
| Provisioning | Do customers need automated deprovisioning? | Whether SSO is a login feature or a user life cycle feature |
| Outside reviews | Who reviews the work, and how long do they usually take? | Calendar dates and when engineers are idle |
How Should You Express Confidence in an Estimate?
Express confidence as a range with the conditions attached, and say what would move it. A three-point estimate (an optimistic, a most likely and a pessimistic figure for each piece of work) is a simple way to build the range from the bottom up, and the spread between the figures shows which pieces carry the most uncertainty.
Check the bottom-up figure against the outside view. Reference class forecasting, which grew out of Daniel Kahneman and Amos Tversky’s research on judgment, starts from how a class of similar past projects actually turned out, then adjusts for what is different about this one. A team estimating its next legacy migration should look at how long its last migrations took against their original estimates, which is usually more sobering than any feature-by-feature build-up.
For a CFO, the presentation matters as much as the method. Effort can be stated in engineer-weeks and converted to cost by finance, and each figure should come with the assumption it depends on:
| Element | What to Tell Finance |
|---|---|
| Range | The effort and duration range for the current phase, and a wider one for the whole program |
| Assumptions | The three or four facts the range depends on, such as the number of active reports |
| What would widen it | The finding that would push the project toward the high end, such as reports that do not reconcile |
| Next narrowing point | The milestone or spike after which the range will be re-forecast |
| Decision needed | What finance must approve now, often only the next phase |
Funding the next phase and committing to the range for the whole program are separate decisions. Approving discovery and spikes first lets the CFO buy information cheaply before committing the larger sum.
How Do You Re-Estimate as the Project Learns?
Re-estimate at defined points, such as the end of a spike, a milestone or the discovery of a broken assumption, and say plainly what changed and why. Track actual effort against the estimate by piece of work, because a variance in one report or one identity provider usually predicts the variance in the ones still ahead.
Decide in advance which evidence would change the forecast. Gun.io’s essay What Would Change Your Mind? draws the distinction between updating an explanation and updating a belief, and the same discipline applies to estimates. If the bank’s parallel run shows totals that do not match, the honest response is to revise the forecast for every remaining regulatory report, not to explain why this report was special. If the first enterprise customer turns out to need identity-provider-initiated login, the SSO estimate changes, and finance should hear it at the next checkpoint rather than at the deadline.
What Moves the Cost Inside an Estimate?
Cost follows the complexity of the problem, the time it takes to solve and the experience the work requires. Experience shows up in the estimate itself: an engineer who has implemented SSO before knows to ask about deprovisioning and certificate rotation in the first week, which narrows the range sooner. Identity protocols are a legitimate reason to bring in specialist depth, because the security boundary is the product.
Team shape matters as well. Fewer people mean less coordination, and one or two strong full-stack engineers who own a feature from database to interface usually produce steadier estimates than a larger team split into layers that hand work to each other. Adding people to a late project rarely brings the date back, because the new engineers need context from the people already doing the work.
How Gun.io Approaches Estimation on a Project
At Gun.io, we start by understanding the scope and complexity of the problem before determining who can solve it and under what engagement model, and there is no minimum commitment. For a legacy system whose behavior nobody fully knows, that can begin with a technical due diligence review of the codebase and data before anyone commits to a full program estimate. In scoped project delivery, Gun.io takes a more extensive role in delivery and engineering management; in staff augmentation, your team directs the work and owns the estimate.
The unknowns in legacy work are real. One of our briefs describes building a finance and operations layer around a legacy ERP instead of replacing the core system, and during the commission work the team found missing source data in the existing spreadsheet process. Findings like that belong in the estimate as soon as they appear, and our engineering solutions cover the legacy modernization and systems integration work these projects involve.
Give Finance a Range, Its Assumptions and the Date It Narrows
A trustworthy estimate states what the team knows, what it is still finding out and when the range will tighten, and it changes when the evidence does. Talk to us about the project you need to estimate, and bring the system, the decisions still open and the date finance needs a number by.
Frequently Asked Questions
Should Engineers Estimate in Hours or Story Points for a Budget?
Story points work for a team planning its own sprints, but a budget needs effort in time units that finance can convert to cost. Teams that use points can translate them using their own measured history of points completed per sprint. Without that history, estimate the budget in engineer-days or engineer-weeks directly.
How Much Contingency Should a Software Estimate Include?
There is no universal percentage, because contingency should be sized from the named risks and the width of the remaining range. A project that has finished its spikes needs less than one that has not. Release contingency through a recorded decision when a risk materializes, so the budget shows what it was spent on.
What Is the Planning Fallacy?
The planning fallacy is the tendency to underestimate the time and cost of a planned task even when similar past tasks overran, a pattern Daniel Kahneman and Amos Tversky described. It persists because estimators focus on the specific plan in front of them and overlook how comparable projects actually went. Reference class forecasting is the usual correction.
Does a Fixed-Price Contract Remove Estimation Risk?
A fixed-price contract moves estimation risk to the vendor; it does not remove it. The vendor prices that risk into the quote, and any work outside the written scope returns as a change order. Fixed price fits well-defined work; for a project with open unknowns, funding discovery first and pricing later phases on what it finds is usually cheaper.
Who Should Produce a Software Project Estimate?
The engineers who will do the work should produce the estimate, with a senior reviewer who has delivered similar projects checking it against the outside view. Estimates produced by someone who will not build the system tend to miss the details that drive the range. For a legacy system, include someone who understands how the business uses it today.
How Often Should a Software Estimate Be Updated?
Update a software estimate at each milestone, after each spike and whenever an assumption it depends on proves wrong. A calendar-only cadence misses the moments when new evidence arrives. Keep each revision on record with the reason for it, so finance can see how the forecast moved and why.