AACE CAPEX estimate classes: what "±30%" actually means

Why an owner should care about estimate classes

In shortThe class tells you what sits behind the CAPEX figure someone has just put in front of you: a scaled analogy to somebody else's plant, or a priced take-off from working drawings. The five AACE classes give the owner, the engineer and the bank one vocabulary. A single line saying "Class 3 estimate per AACE 18R-97" answers half the questions about how far the number can be trusted.

In eighteen years in engineering I have worked with a dozen in-house scales: preliminary, refined, directive, bankable. Every company means something of its own by those words, and the moment two companies meet, the confusion starts. Your "refined" CAPEX can be the contractor's "preliminary". AACE International wrote this problem down in 1997 in RP 17R-97, which notes that parties routinely misread the quality of the information an estimate rests on, the methods used, the accuracy that can be expected and the risk that comes with it.

The 18R-97 classification exists to fix exactly that. It does not make the number more accurate. It makes the conversation about accuracy something other than guesswork. In-house scales do not go away, but 18R-97 is what lets them talk to each other.

Five classes, from a rough figure to a bid check

In shortAACE classes count backwards. Class 5 is built on up to 2% project definition and exists to screen concepts out; Class 3 is usually the basis for budget approval; Class 1 checks a price at tender. The lower the number, the more engineering sits under the figure and the narrower the range around it.

My working version of the scale for a mining and metals owner looks like this. It is rebuilt from AACE 18R-97; the definition percentages and the ranges are the canonical ones.

Two things in that table are easy to read past.

The ranges are asymmetric. Class 5 runs from −50% to +100%: the standard allows twice as much room on the upside as on the down. Construction overruns more often than it undershoots, and the classification says so plainly.

Every cell holds a range of ranges. "Low 10-20%, high 10-30%" at Class 3 means that a simple project on proven technology will sit near −10/+10, and a complex one near −20/+30. Which of those applies to your project is not something the table knows. That comes out of risk analysis on the project itself.

Effort scales with the class as well. Preparing a Class 1 estimate costs anywhere from several to a hundred times a Class 5, on the effort indices in 18R-97. A Class 5 figure can take under an hour, and for its own purpose, deciding whether an idea is worth another look, that is a perfectly good tool.

What "±30%" actually means

In short"±30%" is a probability statement, not a promise. It reads: the final cost has roughly a 90% chance of landing between −30% and +30% around an estimate that already includes contingency. Every clause in that sentence changes the meaning of the number more than it looks.

Taken apart, here is what 18R-97 packs into that notation.

1. The range is measured from an estimate that already carries contingency. If contingency was trimmed to get the budget through the approval meeting, the stated range no longer applies.
2. The estimate itself sits in the middle of the distribution. Contingency is normally set at a 50% confidence level, meaning the probability that the final cost will not exceed the figure quoted. At 50% the outcome is equally likely to land above or below. A coin toss, nothing more.
3. The range covers about 90% of outcomes. One project in ten falls outside it, and the standard is written that way on purpose. 18R-97 notes against its classification matrix that in unusual circumstances the variation can exceed the values shown.

There is a fourth point the standard is silent on and practice is not. John Hollmann, co-author of the mining classification 47R-11 and a contributor to 18R-97, compared what project teams expected with what large projects delivered: teams work to about −10/+30% around the base estimate, while reality runs closer to −20/+120% (Estimate Accuracy: Dealing with Reality, AACE Transactions, 2012). Mining shows the same shape in the academic record. Actual CAPEX comes in around 14% above the bankable feasibility figure on average, and roughly half of projects fall outside the ±15% they claimed (Bertisen and Davis, The Engineering Economist, 2008).

So when someone shows me an estimate "accurate to ±30%", my first question is always the same: ±30% of what, and at what confidence level? If the answer is a pause, the range on that estimate is decoration.

If you are not sure what class the figure on your desk belongs to, that is a reasonable thing to spend ten minutes on.

Why two estimates of the same class differ in accuracy

In shortThe class describes how mature the input is. Accuracy follows from that, but differently for different projects. Two estimates of the same class are not equally reliable. The mining classification 47R-11 puts it bluntly: the accuracy range comes out of risk analysis on the specific project and is never given in advance.

Technology drives the spread more than anything else. In 17R-97 its maturity shifts the range by an order of magnitude, a factor of up to ten. A standard galvanizing line copied from a reference plant and the country's first installation of a new process route can hold identically complete document sets, and the honest range on the second will still be several times wider. Hence a paradox the classification acknowledges outright: a Class 5 estimate on a simple project can be more accurate than a Class 3 estimate on a difficult one (AACE RP 47R-11).

The practical consequence for an owner is that two estimates cannot be compared on class alone. The class tells you what base the numbers were built on. How narrow the range is on your particular project only comes out of quantitative risk analysis run over the estimate.

What actually sets the class

In shortOne parameter decides the class: the level of project definition, meaning which engineering deliverables exist and what state they are in. Not the estimator's qualifications, not the hours spent, not how detailed the spreadsheet looks. 18R-97 is explicit that project definition alone determines the class and that the other four characteristics are secondary.

18R-97 carries a maturity matrix for this: a list of deliverables, process flow diagrams, P&IDs, plot plans, equipment lists, each marked started, preliminary or complete. Want Class 3? Then produce complete process flow diagrams and full equipment lists. Without them, no amount of detail in the spreadsheet will make it Class 3.

Two consequences follow, and both save an owner money.

Detail does not raise the class. A thousand-line estimate built off a sketch is still a Class 5 estimate: there is no engineering under the lines, only assumptions. An estimate resting on undeclared assumptions is expert judgement presented as calculation.

Accuracy cannot be bought with effort. Once a project has reached a given level of definition, no estimator can meaningfully improve the accuracy range by doing a better job of estimating (Hollmann, 2012). Accuracy is produced by engineering; the estimator only records it. The only thing that narrows a Class 4 range to a Class 3 range is the next tranche of engineering: surveys, drawings, vendor quotations on the major equipment.

What class does a bank need?

In shortThere is no written rule saying a bank requires class N. There is industry practice, set down in the mining classification 47R-11: full construction financing opens on a Class 3 estimate, a definitive feasibility study. Pre-feasibility corresponds to Class 4. Compliance is checked by the lender's engineer, and there is no regulation behind it.

The mechanics run like this. International mining reporting, NI 43-101 and the CRIRSCO family, defines a feasibility study as one capable of serving as the basis for a financial institution's decision to fund development of a deposit. AACE 47R-11 translates that into estimate terms: a feasibility study for full financing corresponds to Class 3, pre-feasibility to Class 4, scoping studies to Class 5. Banks in turn appoint a lender's engineer who reviews CAPEX, schedule and contingency before financial close. Bankability is the ability to survive that review. A stamp on the cover page decides nothing.

The Russian public sector arrived at a similar construction by its own route. Projects with budget funding must pass a mandatory technology and price audit of the investment justification (RF Government Decree 563 of 12 May 2017). The function is the same as the lender's engineer: an outside check on the number before the money is released.

The statistics explain why the check is set up so strictly. Mining projects financed on a Class 3 basis still overrun by around 14% on average (Bertisen and Davis, 2008). Class 3 does not remove the risk of an overrun. It makes it measurable.

How the classes map onto Russian design stages

In shortThere is no formal correspondence, and that is worth knowing. The Russian estimating system sets contingency by norm and does not work in probability ranges at all. A practical mapping has settled in and can be used, provided you remember that the class is set by the deliverables, and the name of the stage guarantees nothing on its own.

The trap in that table is that "Stage P" does not guarantee Class 3. If the process side is thin, with preliminary flow diagrams and incomplete equipment lists, an estimate built on Stage P documentation is honestly a Class 4 whatever it is called. The reverse holds too: a well developed pre-feasibility package can deliver a genuine Class 3 before the formal stage is reached.

For Kazakh projects the bridge is easier still. Local practice already runs on the international sequence of scoping study, PFS and DFS, and a bankable feasibility study there means a DFS, which is Class 3.

Contingency is not a cushion

In shortBy the AACE definition (10S-90), contingency is included in most estimates and is expected to be spent. It is not insurance against a bad day. It is part of the forecast cost of the project, and set at 50% confidence it is spent in full on average.

Contingency covers what will certainly happen but cannot be named in advance: design refinements, clashes found on site, additional purchases. It does not cover major changes of scope, force majeure, escalation or currency moves. Those have their own lines in the cost structure, including a separate management reserve, which should never be confused with contingency.

That gives you a simple health check on a budget. If contingency was "optimised" at the approval meeting because nobody plans to make mistakes, the budget is already low, and the accuracy range stated in the estimate no longer holds. An optimistic estimate does not make a project cheaper. It moves the correction to a stage where corrections cost more. How contingency is built from risk rather than from a norm is a note of its own.

Five mistakes owners make with estimate classes

In shortEvery expensive mistake with estimate classes comes down to the same story: asking a figure for something its class cannot give. Five that I have met on live projects, with what each looks like from the owner's side and where it ends up.

1. Comparing figures of different classes. Say a contractor comes back at eight billion and the investment justification carried six. That comparison puts a Class 3 against a Class 5 without normalising anything: different scope, different pricing, different contingency. Align the basis first, then compare.
2. Sanctioning construction on a Class 4. Mining practice is to open full financing on a Class 3. Decisions taken on a thinner base get punished statistically: the less definition at the start, the longer the tail of overruns.
3. Asking for the estimate to be "tightened to ±10%" without further engineering. Accuracy is produced by engineering, which has a cost and a duration. Narrowing the range for free means writing a different number in the same cell.
4. Cutting the top off the range. A corporate rule requiring "a feasibility study accurate to ±10%" only forces the risk analysis to arrive at the answer that was ordered. Hollmann treats targets like these as evidence that risk is being ignored: the owner gets the answer they asked for.
5. Writing "Class 3 estimate" into a contract without a deliverables checklist. The class on its own does not define scope of work. Without an agreed list of documents and their status, you are buying a label.

Five questions to ask your estimator

In shortThe five questions below take ten minutes of conversation and reveal the real class and quality of an estimate faster than a week spent inside the spreadsheets. They need no special preparation, only knowing what to ask and which answers should worry you.

1. What class is this estimate, and against which list of deliverables was the class set? Ask for the document list with statuses, not just the number.
2. What sits in the basis: pricing date, escalation, exchange rates, contingency? An accuracy range without a basis means nothing.
3. What is the range measured from, and at what confidence level? A good answer contains the words "after contingency" and "percentile". A bad one is "well, plus or minus thirty".
4. Where does the range come from: the table in the standard, or risk analysis on this project? The table gives the expectation for a class. Only risk analysis gives the spread on yours.
5. Can every number show its source? A rate, a quantity, a vendor quotation. A number whose source cannot be shown does not belong in the model.

You do not have to be a cost engineer to ask these. You only have to stop taking "±30%" as an incantation. Bad news found during a feasibility study is cheaper than the same news found on site. If you want a specific estimate looked at, write to me.