Analytic Hierarchy Process (AHP)
The Analytic Hierarchy Process (AHP) is a multi-criteria decision method developed by Thomas Saaty. It breaks a choice down into a three-level hierarchy, the goal at the top, the criteria in the middle, the options at the bottom, then derives the weight of each criterion from a series of pairwise comparisons: for each pair of criteria, the decision-maker judges which one matters more and by how much, on a scale of 1 to 9. These judgements fill a matrix from which a priority vector summing to 1 is drawn. The method adds a control no direct weighting offers, the consistency ratio, which measures how far the judgements contradict one another and signals when they need to be revised. The deliverable is a ranking of options backed by traceable weights and a proof of consistency.
Goal
The Analytic Hierarchy Process answers a precise class of decisions: choosing among several options judged on several criteria that pull in opposite directions, when no single figure, neither price nor score, is enough to settle the matter. Cost against functional coverage against risk against strategic fit: the decision-maker senses that one criterion weighs more than another without being able to say by how much, and it is that by how much the AHP produces. The BABOK places the technique among the decision-analysis approaches (§10.16), alongside the decision matrix and pro versus con analysis, as a tool for choosing under multiple criteria.
The method does three things a direct weighting does not. It decomposes the problem into a goal-criteria-options hierarchy, which isolates each judgement. It derives the criterion weights from pairwise comparisons, because a human being compares two elements more reliably than assigning a percentage to seven at once. It audits those judgements through a consistency ratio, a measure of their internal contradiction. The deliverable is a ranked set of options, backed by defensible criterion weights and by numerical proof that the judgements which produced them hold together.
Usage
When to use it
- High-stakes choice on conflicting, partly intangible criteria: cost, quality, risk and fit that no single amount can separate.
- Contested or fuzzy criterion weights: stakeholders cannot say "40%" but can judge that one criterion clearly outweighs another.
- Decision that must be defensible and auditable: a purchase or a governance trade-off where the reasoning is on show, consistency ratio in support.
- Group decision with divergent priorities: the comparisons are elicited per stakeholder, then aggregated by geometric mean.
When not to use it
- Low-cost or reversible decision: the load of pairwise comparisons is out of proportion, prefer a weighted decision matrix or a pro versus con analysis.
- Many criteria or many options: the number of comparisons explodes and fatigue corrupts the judgements, prefer a multi-criteria decision analysis designed for scale.
- Stakes dominated by uncertainty and chance: AHP weights preferences, it does not model probability, prefer a decision tree under decision analysis.
Description
Structuring the decision as a hierarchy
It all starts with a three-level hierarchy. The goal sits at the top, in a single statement: choose the software, select the vendor, settle on the solution option. The criteria form the middle level, the factors the choice is judged on, and a criterion may itself split into sub-criteria where needed. The options sit at the bottom, the concrete candidates, each to be evaluated against every criterion. This decomposition separates judgements that, mixed together, would contaminate one another, and it makes each comparison local and therefore manageable.
The comparison machinery applies twice within this structure. First to weight the criteria against one another, which gives the relative importance of each criterion. Second, criterion by criterion, to compare the options against one another under that criterion, which gives the relative performance of each option on each. The final ranking combines the two: an option's score is the sum, over all criteria, of its weight under the criterion multiplied by the criterion's weight. The mechanism is most simply illustrated on the weighting of the criteria, the ranking of the options using the same arithmetic.
Comparing pairwise on Saaty's fundamental scale
For each pair of elements at the same level, the decision-maker judges the intensity with which one outweighs the other, on Saaty's fundamental scale of 1 to 9. A 1 states equal importance, a 3 a moderate importance of one over the other, a 5 a strong importance, a 7 a very strong importance, a 9 an extreme importance. The even values, 2, 4, 6 and 8, serve as intermediates when a judgement falls between two notches. These judgements fill a square matrix of dimension n for n elements compared, governed by three rules. The diagonal is 1, every element being equal in importance to itself. The matrix is reciprocal: if cost outweighs support with intensity 5, then support against cost is 1/5, and each cell below the diagonal is the inverse of its mirror. It follows that the decision-maker enters only the cells above the diagonal, that is n(n − 1)/2 comparisons: three criteria call for 3, seven call for 21, ten call for 45. That count alone governs the cost of the technique and its limit of scale.
Deriving the priority weights
The matrix of judgements converts into a priority vector, one weight per element, summing to 1. The exact form is the matrix's principal eigenvector, the eigenvector associated with its largest eigenvalue, and that is what software computes. By hand, a three-step approximation is enough and lights up the logic. Sum each column. Normalise each cell by dividing it by its column sum, so that each column now sums to 1. Average each row of the normalised matrix, and those row averages are the weights sought, which sum to 1. This approximation by average of the normalised columns is exact when the matrix is perfectly consistent and very close to the eigenvector otherwise: it is the right method for an example legible by hand, provided it is announced as an approximation of the eigenvector.
Checking consistency
Human judgements have no reason to be perfectly transitive. If cost is worth twice support and support three times something else, strict consistency would want a ratio of six, but a person may say four or eight. AHP does not forbid it, it measures it. First λmax is computed, the largest eigenvalue: by hand, multiply the matrix by the weight vector, divide each component of the result by the corresponding weight, and average. For a consistent matrix λmax equals n, and it is always greater than or equal to it. From it comes the consistency index, CI = (λmax − n)/(n − 1), then the consistency ratio, CR = CI/RI, where RI is Saaty's random index, the average CI of a large number of randomly filled reciprocal matrices of the same order. Its reference values are 0.58 for n = 3, 0.90 for n = 4, 1.12 for n = 5 and 1.24 for n = 6 (n = 1 and n = 2 are trivially consistent, RI is 0 there and the ratio cannot be computed). The usual threshold is CR ≤ 0.10: below it, the judgements are consistent enough to be kept, above it, the most contradictory judgement must be found and revised.
The pitfalls
What the consistency ratio covers
The ratio measures the internal consistency of the judgements. A matrix fabricated out of nothing can be perfectly consistent and perfectly wrong: a low CR attests the logic of the comparisons; it says nothing about the preferences they express.
Explosion of comparisons
The n(n − 1)/2 comparisons, multiplied by a round of options under each criterion, quickly saturate the decision-maker, and a tired decision-maker fills the matrix at random. This is the technique's limit of scale.
Rank reversal
Adding or removing an option can reorder options that are not concerned, a counter-intuitive effect specific to aggregation-based synthesis. Fix the option set before comparing and be wary of a ranking that flips when the list moves.
False precision
Weights to three decimals drawn from coarse verbal judgements dress an opinion up as a measurement. The figures are only as solid as the comparisons that feed them.
AI considerations
AHP has a purely arithmetic core and an irreducibly human layer, and the boundary between the two says where assistance helps and where it harms. On the arithmetic side, the help is clear. Computing the eigenvector, λmax, the consistency index and ratio is work the machine does without error and instantly, where computing by hand is tedious and fallible. A tool flags on the spot a CR above 0.10, and better, it points to the most inconsistent judgement and proposes the revision that would reduce the ratio most, Saaty's repair logic being algorithmic. On a group decision, it aggregates by geometric mean the matrices of several stakeholders and brings out the pairs where they diverge most. Upstream, a language model proposes a first pass of criteria and sub-criteria from a problem statement, which the analyst then trims.
The human layer cannot be delegated. The pairwise comparisons are the stakeholders' preferences: "by how much does cost matter more than support" is a value judgement that AHP exists to capture from those who own the decision. A comparison fabricated by an AI produces a result that is mathematically clean and grounded in nothing, and the consistency ratio does not detect it, since an invented matrix can be consistent. The choice of the criteria themselves, which ones count and to whom they belong, is the framing decision: handing it to a tool buries a value trade-off inside a mechanism. The machine holds the arithmetic and the diagnostics, it does not supply the priorities.
Examples
A small or mid-sized company in French-speaking Switzerland is choosing a hosted project management tool. Before ranking the vendors, it weights its three criteria against one another: cost, features and local support. The judgements read as follows: cost moderately outweighs features (3) and strongly outweighs support (5), features moderately outweigh support (3). These three judgements above the diagonal are enough to fill the matrix, the diagonal being 1 and the lower cells being the reciprocals.
| Criterion | Cost | Features | Support | Weight |
|---|---|---|---|---|
| Cost | 1 | 3 | 5 | 0.633 |
| Features | 1/3 | 1 | 3 | 0.260 |
| Support | 1/5 | 1/3 | 1 | 0.106 |
The weights read in a sentence: cost dominates the decision, at close to 63%, features come next, at 26%, local support remains a supporting criterion, at 11%. The consistency ratio, at 0.033, holds well below the 0.10 threshold, so the three judgements hold together and the weights are kept as they are. It is these weights that will serve to combine the vendors' scores, each vendor having been compared to the others under each criterion by the same machinery, to produce the final ranking. A CR that had exceeded 0.10 would have sent the decision-maker back to the matrix, to find the comparison that contradicts the other two and re-set it, before any ranking.
Visualisations
Two representations carry the technique, and they show two different things. The hierarchy goal-criteria-options, in linked boxes, shows the decomposition that defines AHP: a single goal at the top, the criteria in the middle, the options at the bottom, each option linked to each criterion. It is the model of the decision, the one that says where the comparisons apply, and it serves as a header illustration as much as a landmark in the body.
The comparison matrix and its weights, in rows and columns, is the artefact the technique produces, and that is why it is rendered as a table: the pairwise judgements, the reciprocals below the diagonal, the column of derived weights and, in the caption, the consistency ratio that attests the judgements hold together. A reviewer recomputes the weights and the ratio from the judgement cells alone, which a table allows and an image would forbid.
Cost
| Phase | Level | Rationale |
|---|---|---|
| Preparation | Medium | Building the hierarchy calls for naming the goal, settling the criteria and sub-criteria and gathering the right stakeholders. The choice of criteria is the real work, and it conditions everything that follows. |
| Execution | Medium to High | The number of comparisons grows as n(n − 1)/2 and repeats for the options under each criterion. Eliciting the judgements, then computing the weights and the consistency ratio, carry the bulk of the effort, a spreadsheet or a dedicated tool holding the arithmetic. |
| Documentation | Low | The matrix, the weights and the consistency ratio are their own record. The file holds as soon as the judgements and their authors are logged, which is also what gives the technique its audit value. |
Tooling
The spreadsheet is enough for a modest hierarchy: one sheet holds the matrix, the column sums, the normalisation, the row average and the computation of λmax, the index and the consistency ratio, and everything recalculates as soon as a judgement changes. It is the tool of the first pass and of the teaching example.
Dedicated AHP software (Super Decisions, published by Saaty's Creative Decisions Foundation, Expert Choice, the AHP extension of the decision-tools family) holds the full hierarchy, computes the exact eigenvector rather than the approximation, flags the most inconsistent judgement and handles the aggregation of several decision-makers. They come into their own as soon as the criteria, the sub-criteria and the options multiply.
The broader multi-criteria decision-analysis and decision-support platforms integrate AHP among other weighting and ranking methods, useful when the organisation wants to compare the AHP result with that of another approach, or when AHP only serves to produce the weights of a larger decision model.
Sources
- IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.16 Decision Analysis: the placing of AHP among the decision-analysis approaches under multiple criteria. The BABOK names the technique and situates it, it does not describe its mechanics, which are Saaty's.
- Thomas L. Saaty, A scaling method for priorities in hierarchical structures, Journal of Mathematical Psychology, 15(3), 1977, 234-281: the founding article, the fundamental scale of 1 to 9, priorities by eigenvector and the consistency measure.
- Thomas L. Saaty, The Analytic Hierarchy Process: Planning, Priority Setting, Resource Allocation, McGraw-Hill, 1980: the work that names and codifies the method, the hierarchy, the pairwise comparisons, the synthesis of priorities and the random index.
- Thomas L. Saaty, Decision making with the analytic hierarchy process, International Journal of Services Sciences, 1(1), 2008, 83-98: the compact and late restatement of the method, hierarchy, comparisons, priorities, consistency and synthesis, and the 0.10 threshold on the consistency ratio.