The Complete Guide to Delivery Confidence

What Delivery Confidence is, and what it is not

Somewhere in every people business there is a field with a colour in it. Green, amber, red. It sits on a slide, or in a portfolio tracker, or beside a deal in a pipeline review, and it is supposed to answer the most important question the business ever asks: can we deliver what we are about to promise? The trouble is that the colour was typed in by a person. It reflects how that person felt about the work on the morning they updated the tracker, and feelings, as a class of evidence, are optimistic. The status is green because green is the colour of not wanting to have a difficult conversation.

Delivery Confidence is the discipline of replacing that colour with something you can defend. It is a composed, evidence-weighted score for whether a commitment can actually be kept, computed from the real signals that determine delivery rather than asserted by whoever owns the tracker. It is one of ONX’s own frameworks, and this guide describes it as ours: a first-party methodology, born in the Operations layer of the platform, for treating deliverability as a measured question with a traceable answer.

A status field is an opinion with a colour. Delivery Confidence is an argument you can take apart.

It helps to define the thing by what it refuses to be. Delivery Confidence is not a RAG status, not optimism, and not a forecast of intentions; it is a graded, decomposable answer to a factual question, held to the standard of evidence you would want if the number had your name on it. Three distinctions do most of the work.

First, it is not a RAG status. A traffic light is a single colour standing in for a hundred underlying realities, and it collapses the moment you ask why. Delivery Confidence decomposes: it is a score out of one hundred that breaks into named dimensions, so you can see that the eighty-two is really a ninety on margin, a ninety-five on capacity, and a fifty-eight on the one dependency that will sink it. A colour hides that structure. A composed score exposes it.

Second, it is not optimism, and it is not pessimism either. It does not have a mood. It reads the evidence, weights each input by how well that input is actually known, and returns a number that moves only when the evidence moves. If the number is uncomfortable, the discomfort is information, not a failure of team spirit. The whole point is to make the difficult conversation happen in a planning meeting rather than in a penalty clause.

Third, it reports two things, not one. It returns a score, how ready the commitment is, and, separately, a confidence level, how well-evidenced that score is. An eighty-two built on fresh, measured data is a very different object from an eighty-two built on three-week-old estimates, and the framework refuses to let those look alike. This second axis, the honest labelling of how much you should trust the first, is the difference between a number and a defensible number. It is also the through-line of everything that follows, and the reason Delivery Confidence belongs to the same family of ideas as ONX’s evidence hierarchy and business confidence. Before we build it up dimension by dimension, it is worth being clear-eyed about the problem it exists to solve, because the evidence on how often delivery actually fails is sobering.

Why delivery fails, and what over-commitment costs

The uncomfortable premise beneath Delivery Confidence is that delivering on the promise is not the normal case. It is the exception, and it has been for as long as anyone has kept records. A business that assumes it will deliver, and only checks whether it did afterwards, is betting against a very large and very stable body of evidence.

Start with the most-cited number in the field, and with the caution it deserves. The Standish Group’s 2015 CHAOS research, drawn from tens of thousands of projects, classified only 29% as successful, with 52% challenged (delivered late, over budget, or under scope) and 19% failed outright.¹ Those figures are famous, and they are also contested: a peer-reviewed critique in IEEE Software argued that Standish’s method, which rests almost entirely on one-sided estimation accuracy, is unreliable enough that the precise percentages should not be treated as gospel.² We cite the CHAOS numbers as directional rather than exact, and the direction is not in doubt: even on the most generous reading, a large share of committed work does not land as promised.

The Project Management Institute puts a price on that. Its 2018 Pulse of the Profession found that 9.9% of every dollar spent on projects is wasted because of poor performance, which it framed memorably as roughly USD 99m lost for every USD 1bn invested; in the same survey, about 31% of projects missed their goals, 43% ran over budget, and 48% finished late.³ These are not small tails on an otherwise reliable process. They are the process.

The deepest evidence comes from Bent Flyvbjerg’s work at Oxford, built on the largest database of its kind. He calls the pattern the iron law of megaprojects: over budget, over time, over and over again. Roughly nine out of ten large projects run over budget, with overruns of fifty per cent in real terms common rather than exceptional.⁴ And when he widened the lens across more than sixteen thousand projects of every size, the base rates were starker still: only 8.5% finished on budget and on time, and a mere 0.5% finished on budget, on time and delivered the benefits that had been promised.⁵ Read that number again. Ninety-nine and a half times out of a hundred, the thing that was sold is not the thing that was delivered.

Only one project in two hundred lands on budget, on time, and delivers what was promised. Confidence, asserted rather than earned, is the constant.⁵

What links these findings is not incompetence. It is the moment of commitment. The overrun is determined not when the work goes wrong but when the promise is made, on optimistic assumptions that nobody was asked to evidence. Flyvbjerg is explicit that the failures are baked in at the front end, in a confidence that was declared rather than composed. This is precisely the failure a people business cannot afford, because in a people business the commitment is the product. A software firm can over-promise and let a good product absorb it; a business process outsourcer that commits to headcount it cannot hire, at a margin the contract will not bear, has no product to hide behind, only the delivery, done at a loss. The full mechanism is set out in our note on the cost of selling what you cannot deliver.

Flyvbjerg attributes the pattern to two forces, and both are worth naming, because Delivery Confidence is designed to counter each. The first is optimism: the honest cognitive bias that leads planners to take an inside view of their own commitment and assume it will go better than the reference class of similar commitments ever has. The second, blunter, is strategic misrepresentation, a commitment made deliberately rosy because a rosy commitment is the one that wins the bid or clears the approval. A composed, evidenced score is a defence against both.⁴ It replaces the inside view with the outside evidence, and it makes a deliberately optimistic input visible as a low-grade, stated assumption rather than a fact, so that neither the delusion nor the deception survives contact with the number.

The lesson is not that people are bad at delivery. It is that optimism at the point of commitment is a structural, measurable, expensive default, and that the antidote is not more effort during delivery but more honesty before it. Delivery Confidence is that honesty, expressed as a number. And the first thing it changes is when you find out.

A leading indicator, not a lagging one

There is a comforting illusion in a wall of green dashboards, and it is worth naming, because it is the single most common reason a business is surprised by a delivery failure it could have seen coming. The illusion is that a healthy dashboard means healthy delivery. It does not. It means delivery was healthy, in the period the dashboard covers, which has already happened.

Almost every operational metric a people business steers by is lagging. Service level, average handle time, first-contact resolution, utilisation, on-time delivery: each is a measurement of work that is already done. They are the scoreboard, and a scoreboard is invaluable for telling you the result of the last play. It is useless for telling you whether the next commitment is deliverable, because that commitment has not been delivered yet, so there is nothing to measure. By the time a lagging metric turns red, the money is spent and the client is unhappy.

Delivery Confidence is deliberately built on the upstream signals that move before the dashboard does, so it can warn you about a failure while it is still preventable rather than after it has been paid for. It reads the leading indicators, the ones that predict the lagging ones. Consider attrition. A contact-centre programme staffed today looks fine on today’s service level, but if attrition is running at the levels the industry now reports, 31.2% in Metrigy’s 2024 data⁷ and, on some measures, above 50%,⁸ then a chunk of the team you are committing will not be there in ninety days. Attrition is a leading indicator of a service-level collapse that has not happened yet, which is exactly why we treat attrition as a leading indicator and fold it into the confidence score rather than waiting for it to show up as missed SLAs.

A green service-level dashboard tells you the promise held last month. It cannot tell you whether the next one will.

The same logic runs through every dimension. Hiring velocity leads staffing, which leads occupancy, which leads service level. Ramp completion leads productivity, which leads quality. The change-request rate leads scope creep, which leads margin erosion. In each case there is an early signal, visible now, that predicts a late outcome, visible only after the commitment is due. A business that reads only the late outcomes is permanently driving by its rear-view mirror. This is the difference we draw between the metrics that predict outcomes and the ones that merely describe them, and it is the whole reason Delivery Confidence is computed at the moment of commitment rather than reported afterwards.

None of this makes the lagging dashboards useless. You still need the scoreboard; you simply must not mistake it for a windscreen. The relationship between the two is where the real value sits: Delivery Confidence makes a prediction before the work, and the lagging metrics later score that prediction, which is how the framework learns whether its own weightings were right. A leading indicator that is never checked against outcomes is just a different kind of guess. One that is scored against reality, every time, becomes an instrument that gets sharper the more the business uses it. To build that instrument, you have to know what it is made of, which means the dimensions.

The dimensions of delivery risk

Deliverability is not one thing. It is the conjunction of several independent things all having to be true at once, and the reason a single status colour is so misleading is that it forces a dozen separate risks through one narrow pipe. Delivery Confidence resists that compression. It keeps the risks apart, names each one, evidences each one on its own terms, and only then composes them. The framework works with about nine dimensions, and it is worth walking each, because the discipline is in the decomposition.

Every dimension is a question the business would rather not ask out loud before it signs, which is precisely why the framework asks it for you. Take them in turn.

Capacity and hiring asks the bluntest question: can we field the seats and hire the heads, in the right place and language, by go-live? It is evidenced by the live workforce plan, seat inventory, and the hiring pipeline’s fill velocity. In a market where professional-services utilisation has slid to 66.4%,⁶ capacity headroom is neither free nor infinite, and a commitment that assumes people who do not yet exist is the most common way delivery fails.

Skills and ramp asks whether those people will be productive by the date, not merely badged. A seat filled is not a seat producing; new hires climb a ramp to competency, and a plan that counts them as fully effective on day one is counting fiction. It is evidenced by training completion and historical ramp curves, and it is where an optimistic staffing model quietly overstates capacity.

Workforce and attrition risk asks whether the team you staff will still be there in ninety days. Given attrition running anywhere from 31.2%⁷ to above 50%⁸ across the industry, a commitment must be evidenced against the tenure mix and attrition trend of the specific team delivering it, not an aspirational retention number.

Dependencies and upstream inputs asks about the things you do not control: client data, client systems, third-party integrations, another team’s deliverable. These are the classic silent killers, because they are usually evidenced only by someone else’s promise. A dependency that arrives after go-live binds the whole commitment regardless of how strong everything else is.

Commercial and margin headroom asks whether there is enough margin to absorb the delivery risk you are taking. A wafer-thin deal has no capacity to survive an overrun; the same slippage that a fat-margin contract shrugs off will push a tight one into loss. With cost reduction still the primary motive for more than 70% of buyers, who now judge providers on fully-loaded cost,¹⁰ margin headroom is under constant pressure, and it is evidenced by the priced cost-to-serve set against the contracted rate.

Client relationship and sponsor risk asks whether the relationship is broad and healthy or narrow and quietly fraying. A single sponsor, a rising escalation count, a silent satisfaction score: these predict churn and dispute long before the service level moves. It is evidenced by sponsor coverage, open escalations and sentiment, and it is where a technically deliverable commitment can still go wrong.

Compliance and governance asks whether you can deliver within the rules that bind this particular work. For people businesses the rules are tightening: the EU AI Act now classes AI used in recruitment, task allocation and performance monitoring as high-risk,¹¹ which means a commitment involving those uses carries obligations that must be evidenced, not assumed. A commitment you cannot deliver lawfully is not deliverable at any price.

Timeline and the binding constraint asks the question that ties the others together: what is the earliest date on which every requirement clears, not the date someone hopes for? It is not a fresh measurement so much as a convergence of the others, the latest of all the clear-by dates, and it is the dimension that most often reveals that a confidently-sold go-live was never real.

Change and scope volatility asks how stable the thing you are committing to actually is. A specification that churns weekly is a commitment to a moving target, and volatility itself is a risk independent of the current scope. It is evidenced by the change-request rate and requirements churn, and it is the dimension that most reliably predicts the margin erosion nobody priced in.

Laid out as a table, the discipline becomes clearer: each dimension has its own question, its own natural evidence, and a characteristic quality of that evidence, which is the subject of the next section.

The exact set flexes by delivery type, a staffing placement weights different dimensions from a multi-year contact-centre programme, and the framework is versioned precisely so the composition can improve as outcomes teach it. But the shape holds: keep the risks apart, evidence each honestly, and let the composition, not a single anxious colour, produce the answer.

Evidence, and the quality of evidence

Here is the mistake that makes most risk scoring worthless: it treats a number someone typed with the same weight as a number the system measured. A confidence score that adds a hard fill-rate from the applicant-tracking system to a hopeful attrition assumption from a spreadsheet, and averages them as equals, has laundered a guess into a fact. Delivery Confidence exists partly to refuse that laundering.

Every input to the score carries not only a value but a grade for how well that value is known, and a low-grade input is never allowed to masquerade as a measured one. This is ONX’s evidence hierarchy applied to deliverability, and it distinguishes at least three kinds of input.

• Measured. A fact drawn live from a system of record: seats actually available in a named site, candidates actually in the pipeline, escalations actually open, cost actually booked. This is the strongest evidence, and it is stamped with where it came from and when it was last refreshed.
• Modelled. A value produced by a defensible calculation rather than a direct observation: weeks-to-fill inferred from historical hiring velocity, ramp-to-competency projected from past cohorts, volume assumed from a forecast. Legitimate, useful, and clearly weaker than a measurement, so it is labelled as modelled and never dressed as fact.
• Stated. A value that rests on someone’s assertion: a client’s promise that data will arrive on the first, a sales note that the sponsor is engaged, an assumption entered by hand. This is the weakest grade, and the framework treats it accordingly, because most delivery failures are not made on bad data; they are made on assumptions wearing the costume of data.

Two consequences follow, and both are deliberate. The first is that each dimension shows its provenance: alongside its score sits the source that produced it and the date it was last updated, so a reviewer can see at a glance that the capacity number is live and the attrition number is three weeks stale. Freshness is itself a form of evidence quality; a measurement from a month ago is closer to a modelled value than a measured one, and the framework marks it so.

When a measurement becomes an assumption

Evidence does not stay fresh. A seat count pulled live this morning is a measurement; the same figure a fortnight later, unrefreshed, is closer to an assumption, because the world has moved and the number has not. Delivery Confidence treats the age of an input as part of its quality, decaying a stale measurement toward the weight of a modelled or merely stated value rather than pretending it is still hard fact. In practice a dimension can slip from green to amber without anyone touching the underlying commitment, simply because the evidence beneath it went out of date. That is precisely the warning a business wants: the promise did not change, but the right to be confident about it did. The remedy is not to distrust all data but to refresh the inputs that matter before the commitment is made, not after it has failed.

The second consequence is the confidence level introduced at the start. Because every input is graded, the framework can compute not just a score but how much of that score rests on measurement versus assumption, and it reports that as a separate high, medium or low confidence. This is why two commitments can both score eighty-two and mean entirely different things: one assembled from fresh, measured inputs and worthy of the number, the other propped up by stale estimates and flagged as such. A single colour cannot carry that distinction. A graded, sourced, dated composition can.

Most delivery failures are not made on bad data. They are made on assumptions wearing the costume of data.

The discipline this imposes is uncomfortable at first and liberating afterwards. It forces a business to admit, in the moment of commitment, how much of its confidence is actually earned. Often the honest answer is: less than the slide suggested. But a low-confidence eighty-two is a solvable problem, you know exactly which assumptions to go and verify to raise it, whereas a green status hides the very fact that it was built on air. Evidence quality turns a vague unease into a task list. And it sets up the harder question that the composition itself has to answer, which is what to do when the dimensions disagree.

Why the worst binding dimension decides

Suppose you have scored all nine dimensions honestly, each with its evidence and its grade. Now you have to combine them into a single answer, and here lies the temptation that quietly ruins most scorecards: to average. Averaging feels fair and looks rigorous, and it is exactly wrong, because a business does not experience the average of its constraints. It experiences the one that fails.

Consider a commitment, drawn as an illustration rather than from any client: twelve German-speaking agents, go-live in September. Sales confidence is high. The workforce plan is solid. The margin clears the floor. Four dimensions score in the nineties. But the platform integration the service depends on is not safe until nineteen days after the contractual go-live. Average the five and you get something around seventy-six, which looks committable, a comfortable amber shading to green. Ask the only question that matters, can we keep this promise on this date, and the answer is an unambiguous no. The technology dependency binds, and no amount of strength elsewhere compensates.

A commitment that scores ninety in four dimensions and twenty in the fifth is not seventy-six per cent healthy. It is blocked, and the framework says so.

Delivery Confidence is governed by its worst binding dimension, not its flattering average, because the binding constraint is the thing the business will actually hit. This is the Theory of Constraints applied to commitment: a system’s throughput is set by its bottleneck, and dressing up the non-bottlenecks changes nothing. In the framework, a dimension in the red does not get diluted by the greens around it; it caps the recommendation and is named as the constraint, so the conversation moves from an argument about a blended number to a specific, actionable fact: this commitment is blocked by technology readiness, and here is the date it clears. The full reasoning lives in our guide on why the worst constraint decides.

The practical effect is that recommendations become dates and constraints, not scores. Convergence, the act of taking every dimension’s clear-by date and committing only to the earliest date that clears them all, turns a vague seventy-six into a usable instruction: commit for the nineteenth, or move the dependency, or descope. One late dimension moves the whole answer, which is not pessimism but an accurate description of how delivery behaves. It also explains why a single unvalidated assumption can, and should, hold everything: if a launch-critical input is merely stated rather than measured, it binds, because the alternative is discovering it in production with a penalty clause attached.

This is where the two axes from earlier meet. The worst dimension can bind on its score, it is genuinely red, or on its evidence, it might be fine but nobody has checked. Both are constraints. A commitment held up by an unverified assumption is not safe merely because the assumption is probably true; it is unsafe until the assumption is evidenced, and the framework treats low confidence on a critical dimension as its own kind of amber. The worst constraint deciding is uncomfortable exactly once, before you commit. Averages are comfortable twice: before you commit, and then never again. Which raises the question of what happens after the first score, because a commitment is not a photograph. It is a process, and the number has to move with it.

A score that moves as evidence lands

A Delivery Confidence score is not a verdict handed down once and carved into the deal. It is a living reading of a commitment whose underlying reality changes by the day, and treating it as a one-time gate is a way of throwing away most of its value. The number should move, and the direction it moves in is itself information.

The right way to run Delivery Confidence is as a recomputed value that rises as assumptions are verified and falls as risks materialise, never as a figure asserted once and defended thereafter.When a stated assumption becomes a measured fact, the client confirms the data date, the first cohort actually completes ramp, the integration passes its test, the evidence quality on that dimension improves, and the score and its confidence level should both climb, honestly, because the commitment genuinely became safer. Equally, when attrition ticks up or a dependency slips, the number should fall while there is still time to respond. A score that only ever moves at review time is a score nobody is really using.

This is why the framework bands the score rather than fetishising the exact figure. Above a green threshold the commitment is safe to make; a middle band means proceed with named risks; below a lower threshold it is not yet committable. The bands matter more than the decimals, because the job of the number is to trigger a decision, not to be admired. And crucially, a red or amber score is not a verdict of doom; it is the opening of a task list. The most useful thing the framework produces is not the score at all but the path to green: the specific, ordered set of moves, verify this assumption, source seats from that alternative site, descope this dependency, that would lift the commitment into the safe band, and the constraint each move clears.

To make that concrete, and again as an illustration rather than a client case: a commitment scores seventy-one, an amber, held there by a capacity dimension in the red because the named delivery site is forty seats short. The path to green offers two ordered moves. The first, source the shortfall from an alternative site that has the seats and the language, clears the constraint immediately and lifts the score into the green band; the second, hold the original site and let the hiring pipeline fill the gap, clears it too but adds six weeks. The business now has a real choice with real trade-offs, cost against time, rather than a binary approval, and it can see precisely which single fact, the seats, was deciding the whole commitment. A status colour would have shown none of this. It would have shown amber, and left the reason, and the remedy, for someone to reconstruct later, usually too late.

That reframing changes the whole conversation around a difficult commitment. Instead of a binary, take it or leave it, the business gets a route: here is why it is amber, here is the one dimension holding it there, and here are the two things that would make it green and roughly how long each takes. A commitment that was undeliverable on Monday can be made deliverable by Thursday, on purpose, because the framework told you exactly which lever to pull. This is what it means to treat a decision as something you can work on rather than merely accept or reject, and it is the mechanism by which Delivery Confidence improves outcomes rather than just predicting them.

Because the score moves, it also becomes learnable. Each commitment leaves a trail, the score at signing, how it moved, the constraint that bound it, and, eventually, whether the commitment delivered. Scored against the lagging outcomes that arrive later, that trail tells the business whether its own weightings were right, whether a dimension it treated as minor keeps turning out to be fatal, whether its ramp model is systematically optimistic. A leading indicator that is never reconciled with reality drifts into superstition. One that is reconciled every time compounds into judgement. That reconciliation, the sell scored against the deliver, is where Delivery Confidence stops being an Operations tool and becomes the connective tissue of the whole business.

Connecting the sell to the deliver

The most expensive gap in a people business is the one between the room where a deal is sold and the room where it is delivered. They are usually different rooms, with different systems, different incentives and different definitions of a good outcome, and the commitment made in the first is handed to the second as a fait accompli. Delivery Confidence exists, more than anything else, to close that gap.

A sales team judges a deal on what the pipeline can see: stage, value, probability, close date. But in a people business the most important question about a deal lives entirely outside the pipeline, can we actually hire and deliver this, at the margin we are quoting, on the date we are promising? A forecast that ignores delivery reality is a forecast of intentions, not outcomes. Win rate, the metric most sales functions steer by, is on its own a lagging indicator that describes what already closed, not what can be kept, which is why we treat win rate as a lagging indicator rather than a target. Even a disciplined proposal function, choosing its bids carefully and running a formal process, wins only around 43 to 45% of what it enters,⁹ and the bids worth winning are the deliverable ones, a judgement the pipeline cannot make alone.

Delivery Confidence puts a deliverability score on the deal before it is committed, so the business qualifies against delivery reality rather than against optimism. Run early, in the deal room, it answers the question the pipeline cannot: not how likely are we to win this, but should we, given what it would take to deliver it. A deal that scores red on capacity is not a deal to chase harder; it is a deal to requalify, renegotiate, or route to a site that can actually staff it. This is the practice we describe as qualifying against delivery reality, and it is the single most valuable use of the framework, because the cheapest delivery failure to fix is the one you decline to sell.

The mechanism that makes this work is a shared spine beneath the functions. When the deal is scored, the capacity dimension reads the real workforce plan; the hiring dimension reads the actual pipeline and, where there is a gap, can open a hiring requirement against it; the margin dimension reads the priced cost-to-serve from finance. The commitment made in sales is therefore evidenced by the functions that will have to keep it, which is the essence of what we call cross-module intelligence. Sell, hire and deliver stop being separate arguments and become one composed judgement. It is also why workforce planning in a people business should not begin in operations after the deal lands, but in sales before it does.

The same discipline applies after the win, not only before it. A renewal is a commitment too, and the most dangerous kind, because it arrives looking like a formality. An account whose service levels are green can still be quietly undeliverable at the renewed rate if its true cost has crept up, its sponsor has narrowed to a single person, and its recent change requests were scoped below cost. Running Delivery Confidence at renewal, rather than treating the date as an event that simply happens, catches the account that is about to be re-committed at a rate it can no longer bear. This is part of what we mean by attending to what happens after a deal is won, and it is why workforce planning should start in sales rather than waiting for delivery to inherit a promise it was never consulted on.

Close the loop and the economics change. The deals that get committed are the deals that can be delivered at the quoted margin; the ones that cannot are caught while declining them is still cheap. Given base rates where nine in ten large commitments overrun⁴ and utilisation is already thin at 66.4%,⁶ shifting even a fraction of undeliverable commitments out of the book before they are signed is worth more than any efficiency gain during delivery. The sell and the deliver, scored against each other, become a single system that gets better at choosing what to promise. Which leaves the practical question: how does a business actually start running this?

How to operationalise Delivery Confidence

A framework that lives only in a report changes nothing. Delivery Confidence earns its keep when it becomes a habit, a thing the business does at specific moments, with specific evidence, as a matter of routine rather than heroics. The good news is that adopting it is a sequence of small, unglamorous moves, none of which requires a transformation programme to begin.

Operationalising Delivery Confidence is less a software project than a change in when and how a commitment is allowed to be made, and it can start with a single dimension on a single deal. The sequence that works looks roughly like this.

1. Score the commitment before it is made, not after. Put the confidence check at the gate: before a deal is committed, before a go-live is promised, before a renewal is signed at the old rate. The moment matters more than the mechanism. A commitment nobody scored is a commitment nobody interrogated.
2. Decompose, and name the dimensions out loud. Even in a spreadsheet on Monday, stop reporting one colour and start reporting the nine questions. The act of separating capacity from margin from dependency from compliance is most of the value, because it stops a strong dimension from hiding a fatal one.
3. Grade the evidence honestly. Beside each dimension, mark whether its input is measured, modelled or merely stated, and how fresh it is. This is the discipline that separates a defensible score from a hopeful one, and it is available to any team willing to be candid about what it actually knows.
4. Let the worst binding dimension govern the recommendation. Resist the urge to average. If one dimension is red on score or red on evidence, the commitment is not green, and the honest output is the constraint and its clear-by date, not a blended number that lets everyone feel better.
5. Recompute as evidence lands, and keep the trail. Re-score when assumptions are verified or risks appear, run the path to green when a commitment is amber, and, when the work is finally delivered, record whether the prediction held. That reconciliation is what turns the score from a guess into an instrument.

The first two steps you can do this week with the systems you already own, and they will surface at least one commitment you were about to make that you should not. The later steps are where the technology helps, because grading evidence at source, reading the live workforce plan, refreshing the pipeline, converging the clear-by dates, is exactly the kind of thing that decays into stale spreadsheets if a person has to do it by hand. A shared spine on which each function publishes its facts, with their evidence quality, is what makes the habit sustainable at the scale of a whole portfolio rather than a heroic effort on one flagship deal. That is the architecture behind the ONX platform, and the deliverability check is one of the decisions it is built to run in a Decision Room.

But the technology is downstream of the discipline. The habit, scoring the commitment before you make it, decomposing the risk, grading the evidence, honouring the constraint, and learning from the outcome, is available to any organisation willing to stop typing a colour into a field and start composing a number it can defend. What that discipline is really an instance of is worth naming, because it locates Delivery Confidence inside a larger idea.

Where Delivery Confidence goes wrong in practice

A framework that measures deliverability can be misused in ways that quietly return a business to exactly the false comfort it was meant to remove. The failure modes are worth cataloguing, because they are predictable, and because avoiding them is most of what separates a Delivery Confidence practice that changes decisions from one that merely decorates them.

The commonest way to ruin a deliverability score is to turn it from a measurement into a target, at which point people optimise the number instead of the delivery. This is Goodhart’s law, and it arrives the moment a score becomes a gate that a deal must clear to win approval. Faced with a threshold, the rational move is to make the number look right rather than make the commitment safe: to enter a hopeful fill rate, to grade a stated assumption as though it were measured, to quietly average the red dimension back into the greens. The defence is built into the framework rather than bolted on. Because every input carries its source and its evidence grade, a score inflated by optimistic inputs shows up as high on the surface and low in confidence underneath, and the audit trail records who entered what. A number you can inspect is much harder to game than a colour you can simply type.

A related failure is reintroducing the average through the back door. The binding-constraint rule is uncomfortable precisely when it is most useful, when four dimensions are strong and one is fatal, and the temptation to say it is only one red, surely the overall picture is fine, is strong. That one red is not a blemish on an otherwise good commitment; it is the commitment. Every time a business overrides the constraint because the blended number felt reassuring, it is choosing the comfortable answer over the true one, and paying for the difference in delivery.

The third failure is treating stale evidence as fresh. A score is only as current as the inputs beneath it, and inputs decay: last quarter’s attrition, a pipeline snapshot from six weeks ago, a margin figure taken before the scope grew. The framework marks freshness for exactly this reason, but a marking can be ignored, and a business that lets old numbers stand in for current reality has rebuilt the very illusion of a green dashboard it set out to escape. Confidence in a commitment should degrade as its evidence ages, and a practice that never refreshes its inputs is trusting a photograph of a moving thing.

The fourth is scoring once and filing the result, the one-time gate again, wearing a different hat. A score that is computed at the deal and never revisited cannot warn anyone about the risk that appeared after signing, which is where most delivery risk actually lives. The fifth is subtler and more corrosive: using a low score as a weapon. If the person who surfaces an amber is punished for it, people stop surfacing ambers, and the framework goes dark exactly when it is most valuable. The score is a judgement about a commitment, not about the person who reported it, and a culture that confuses the two loses the instrument. Finally, there is over-precision, the long argument about whether a commitment is an eighty-two or an eighty-four, when the band and the binding constraint are the only things that carry a decision. The decimals invite exactly the false rigour the whole discipline exists to replace.

What unites these failures is a slide back toward the single, unaccountable number the framework was built to abolish. Each of them, gaming, averaging, staleness, one-time scoring, blame, false precision, is a way of keeping the appearance of Delivery Confidence while discarding the honesty that made it worth having. The discipline survives contact with a real organisation only if the two axes are read together, the constraint is honoured even when it is inconvenient, and the score is allowed to be uncomfortable. Held to that standard, it is not a scorecard at all. It is a habit of telling yourself the truth about a promise, on schedule, in public.

The operational expression of business confidence

Step back from delivery specifically and the pattern is general. The same three moves, compose the answer rather than assert it, grade the evidence rather than launder it, let the binding constraint decide rather than the average, apply to every consequential judgement a business makes, not just to whether it can deliver. Delivery Confidence is the delivery-shaped instance of a wider discipline.

Delivery Confidence is the operational expression, in the Operations layer, of ONX’s broader idea of business confidence: a decision treated as a governed object rather than a colour or a hunch. Where business confidence describes the general principle, a confidence figure computed from published evidence and weighted by its quality, not typed in by whoever is most senior in the room, Delivery Confidence is that principle pointed at one specific and high-stakes question. The evidence hierarchy that grades measured against modelled against stated is the same hierarchy. The worst-constraint rule that lets one red dimension govern is the same rule. The scored outcome that teaches the next commitment is the same loop. Delivery is simply where the idea bites first and hardest, because in a people business the commitment is the product.

Seen this way, Delivery Confidence is one reading on an instrument that also measures whether a deal is worth pursuing, whether an account should be renewed, whether a market is worth entering. Each is a decision that deserves a composed, evidenced, constraint-aware answer rather than a status colour, and each gets sharper when its predictions are scored against what actually happened. The wider account of that discipline, and why it is emerging now across BPOs, consultancies, MSPs and staffing firms, is set out in our companion report, The State of Decision Intelligence for People Businesses 2026, and the underlying philosophy in the pillar essay on enterprise decision intelligence.

The reason to start with delivery is that it is where the money leaks most visibly and where the evidence of failure is least in dispute. The research is unusually consistent: most committed work does not land as promised,¹ the cost of that is measured in the high single digits of every dollar,³and only a fraction of a per cent of large commitments deliver in full.⁵ A business that learns to score its deliverability honestly before it commits has, in effect, learned the whole discipline on the problem where it pays back fastest. From there, the same instrument extends to every other decision that currently hides behind a colour.

Delivery Confidence, then, is not a dashboard and not a gate. It is a way of being honest with yourself about a promise before you make it, and of getting measurably better at that honesty every time reality settles the question. The businesses that win the next decade in this industry will not be the ones that deliver flawlessly, nobody does that. They will be the ones that stopped promising what they could not keep, because they finally had a number that told them, in time, which promises those were.

Methodology & a note on honesty

A guide about scoring commitments honestly has no licence to be dishonest about its own claims. So a word on how this one was assembled, and on what it deliberately does and does not assert.

Delivery Confidence itself is ONX’s own framework, a concept from the Operations layer of our platform, and it is presented here as ours: a first-party methodology, not an external standard and not borrowed research. The nine dimensions, the two axes of score and evidence-confidence, the worst-constraint rule and the path-to-green mechanic are our design. Where this guide leaves the framework and reaches for facts about the world, why delivery fails, what over-commitment costs, how the industry is changing, every one of those figures is drawn from a named, dated, public source and cited in the References below.

We have been careful with the most-quoted of those figures. The Standish Group’s CHAOS success rates are the best-known numbers in the field and also among the most criticised; we cite them alongside the peer-reviewed critique that questions their method,² and we use them as directional evidence of a well-attested pattern rather than as precise measurements. Where sources disagree, as they plainly do on contact-centre attrition, which ranges from around 31%⁷ to above 50%⁸ depending on definition and sample, we give the range rather than the most alarming point. Bent Flyvbjerg’s base rates are quoted from his published Oxford research and book;⁴⁵ the interview cited for the 8.5% and 0.5% figures reproduces the dataset’s own numbers.

A guide about scoring commitments honestly has no licence to be dishonest about its own claims.

Two things in this guide are explicitly illustrative rather than measured, and are labelled as such in the text. The twelve-agent September commitment used to explain the binding constraint is a worked example, not a client engagement; it exists to show the shape of a mechanism, not to report an outcome. And the specific dimension weightings and band thresholds are described in general terms because they are versioned and vary by delivery type, not fixed constants we are asserting as universal. We have not invented customers, quotes or results, and where a widely-repeated figure could not be traced to a primary source we have left it out: the frequently-cited claim that it costs a fixed amount to replace a departed agent, for example, is quoted across the industry but hard to pin to an original study, so it does no load-bearing work here.

This guide will be revised as the evidence moves, particularly the project-performance research and the regulatory picture around AI in employment, both of which are changing. If you find a claim here you believe is wrong, or a source that has been superseded, we want to know. A reference work earns its standing by being corrected, not by being defended, which is, not coincidentally, exactly the standard Delivery Confidence holds a commitment to.

References

Every figure in this report is sourced below. Third-party links open in a new tab. Where a figure is modelled rather than measured, or where a widely-quoted statistic could not be traced to its primary source, the report says so in the text.

1. CHAOS Report 2015 · The Standish Group, 2015. Of the projects studied, 29% were classified successful, 52% challenged (late, over budget, or under scope) and 19% failed outright. The 2015 edition redefined success around perceived value, which lowered reported success rates.
2. The Rise and Fall of the Chaos Report Figures, IEEE Software 27(1), 30-36 · Eveleens & Verhoef, 2010. A peer-reviewed critique arguing the Standish figures are unreliable because their definitions rest solely on one-sided estimation accuracy; cited here so the CHAOS numbers are used as directional, not exact.
3. Pulse of the Profession 2018: Success in Disruptive Times · Project Management Institute (PMI), 2018. PMI reported that 9.9% of every dollar is wasted due to poor project performance, roughly USD 99m for every USD 1bn invested; and that about 31% of projects miss their goals, 43% run over budget and 48% finish late.
4. What You Should Know About Megaprojects and Why: An Overview, Project Management Journal 45(2), 6-19 · Bent Flyvbjerg (University of Oxford), 2014. Sets out the “iron law of megaprojects” (over budget, over time, over and over again) and documents that roughly nine out of ten megaprojects run over budget, with overruns of 50% in real terms common.
5. How Big Things Get Done (base rates from a database of 16,000+ projects) · Flyvbjerg & Gardner, 2023. From the Oxford project database: only 8.5% of projects finish on budget and on time, and just 0.5% finish on budget, on time and deliver the promised benefits. Figures quoted by Flyvbjerg in interview; the underlying dataset is the book’s appendix.
6. Professional Services Maturity Benchmark 2025 · SPI Research, 2025. Average billable utilisation fell to 66.4% in 2025, the lowest in the survey’s history; high-performing firms reach around 81.2%. Used here as an indicator of delivery capacity headroom.
7. What Metrigy’s Latest AI Data Reveals About Contact Center Staffing · Metrigy, 2024. Contact-centre agent turnover rose to 31.2% in 2024, from 28.1% (2023) and 21.8% (2022).
8. Global Contact Center Survey · Deloitte Digital, 2023. Reported average annual agent attrition of around 52%, illustrating how widely attrition estimates range across the industry.
9. RFP Response Trends and Benchmarks · Loopio, 2025. Average RFP win rates cluster around 43 to 45%, even among firms running a formal proposal process.
10. 2024 Global Outsourcing Survey · Deloitte, 2024. Cost reduction remains the primary outsourcing driver for more than 70% of organisations, and buyers increasingly judge providers on fully-loaded cost rather than headline rate.
11. Annex III: High-Risk AI Systems (Regulation (EU) 2024/1689) · EU Artificial Intelligence Act, 2024. Classifies AI used in employment, worker management and access to self-employment (recruitment, screening, task allocation, performance monitoring) as high-risk, triggering governance obligations.