The State of Decision Intelligence for People Businesses 2026

The people business, defined

There is a category of company the software industry has never quite known what to do with. Its product is not a widget or a subscription; it is people, doing skilled work on someone else’s behalf. A business process outsourcer answering a bank’s customers. A consultancy staffing a transformation programme. A managed service provider keeping a hospital’s systems alive at three in the morning. A staffing firm placing ten thousand workers a month. We will call them, together, people businesses, because the thing they sell and the thing that costs them money are the same thing: human effort, organised well.

They are not a niche. Taken together they are one of the largest employers on earth and, increasingly, one of the more contested corners of the economy. The global market for business process outsourcing alone was worth around USD 302.6bn in 2024 and is forecast to reach USD 525.23bn by 2030, a compound growth rate of 9.8%.¹ The staffing industry is larger still, at roughly USD 620bn in 2024, with the United States (~USD 184bn) and the United Kingdom (~USD 50.8bn) among its biggest single markets.² Management consulting, on the narrower definition used by the specialists who track it, is heading toward USD 110bn, though it grew only 2.9% in 2024, its slowest in years.³ Managed services, blending technology and labour, was around USD 297bn in 2024 and is the fastest-growing of the four, compounding at roughly 15%.⁴

What unites them is not their size but their shape. A software company’s marginal cost of serving one more customer is close to zero; a people business pays, in full, for every hour of every delivery. Its economics are therefore not a function of code but of judgement, made continuously and under pressure: which deals to chase, which to walk away from, who to hire and when, how to staff a shift, whether a contract can be delivered at the price it was sold. Get those judgements right and a people business earns a respectable margin on enormous revenue. Get them wrong, in aggregate and quietly, and the margin disappears while the revenue stays, which is the worse of the two failures because it looks like success.

This report is about those judgements: the decisions on which people businesses live or die, the forces that are making them harder in 2026, and the discipline emerging to make them better. That discipline has a name, decision intelligence, and a growing body of evidence behind it. But to understand why it matters now, you have to start with the number that is quietly moving in the wrong direction across the whole category: the margin.

The margin squeeze

The defining financial fact of the people business in 2026 is that revenue is still growing while profitability is not. The demand is real, the backlogs are healthy, and yet the operating margin, the percentage of each pound of revenue that survives to the bottom line, is under sustained pressure across the listed players who have to disclose it.

The clearest single signal came from Concentrix, one of the two giants of customer-experience outsourcing. For its 2025 financial year it reported a non-GAAP operating margin of 12.8%, down from 13.7% a year earlier, a fall of nearly a full percentage point; a large goodwill impairment pushed its reported GAAP operating margin negative for the period.⁵ Its rival Teleperformance, larger again, held its recurring EBITA margin at 15.0% but guided to only marginal improvement, in effect flat.⁶The market read the combination as a category signal rather than a company one: Teleperformance’s shares fell around 13% in the wake of Concentrix’s results, wiping out a chunk of value on fears about where sector margins are heading.⁷

Revenue is growing while the money per unit of work is not. That is the whole story of the people business in 2026.

The pressure has three sources, and they compound. The first is pricing. Buyers of outsourced work have become far more sophisticated about what they are paying for. Deloitte’s Global Outsourcing Survey finds cost reduction is still the primary motivation for more than 70% of buyers, but they now evaluate providers on fully-loaded cost, management overhead, technology, training and the hidden cost of attrition, rather than a simple price per hour or per seat.⁸That transparency squeezes the provider’s room to hold margin.

The second is wages. In a business where labour is the dominant cost, any gap between what the market pays workers and what long contracts allow the provider to charge falls straight onto the margin. Multi-year deals, particularly rate-capped ones in the public sector, cannot reprice as fast as specialist salaries rise, and the difference is absorbed by the provider. This is the arithmetic of a business whose cost base is people: there is no automation dividend hiding in the model to cushion a wage shock, only the margin.

The third is expectation. As artificial intelligence has entered the conversation, buyers have begun to assume that AI-driven efficiency will lower their bill, and to demand that the savings be passed through in the next negotiation. The provider is thus asked to invest in the technology and then hand the benefit to the client, a squeeze from both ends. We will return to what AI actually does to this model, because it is more disruptive than a pricing conversation. For now the point is simpler: the margin is thin, contested, and moving the wrong way, and it is made of thousands of individual decisions about what to sell, who to hire and how to deliver. That is exactly why the quality of those decisions has stopped being a soft topic and become a financial one, which is the subject of the next section, and the heart of this report. The dynamics here are explored further in our note on why BPO margins keep shrinking.

The decision that never gets stored

Here is the strange thing about a people business, and about almost every business. It records, in meticulous detail, the consequences of its decisions, and almost nothing about the decisions themselves. The CRM holds the deal. The ATS holds the hire. The workforce system holds the roster. The finance system holds the invoice. What none of them hold is the reasoning: what was known at the time, which options were considered, which constraint was expected to bind, who chose, and what they were betting would happen. It keeps the consequences and discards the reasoning, and you cannot learn from a decision you never recorded.

The organisation keeps the outcome and throws away the thinking.

The cost of this is not abstract. McKinsey’s research on organisational decision making found that only about 20% of executives believe their organisations are good at it, and that a striking 61% say most of the time they spend making decisions is used ineffectively.⁹Executives, on average, spend a large share of their working week making decisions, and more than half of that time is felt to be wasted. Scaled up, McKinsey estimated the drag at a typical Fortune 500 company at around 530,000 days of managers’ time a year, roughly USD 250m in salary spent on decision making that the decision makers themselves consider inefficient.⁹

This is not merely wasteful; it is a competitive variable. In a companion study, McKinsey found that speed and quality in decision making go together rather than trading off, respondents who described their decisions as fast were nearly twice as likely to describe them as high quality, and the organisations that combined both were roughly twice as likely to report top-quartile financial returns from their recent decisions.¹⁰Bain’s decade-long work on the same question is blunter still: across a thousand companies it found a correlation of over 95% between how well a company makes and executes decisions and its financial performance, with the best decision makers delivering around six percentage points more total shareholder return than the rest.¹¹

Now place that finding inside a people business, where the product is a sequence of decisions. A software company can make a mediocre decision and let a brilliant product carry it. A BPO cannot; if it qualifies the wrong deal, or commits to a headcount it cannot hire, or renews a contract it is quietly losing money on, there is no product to absorb the mistake, only the delivery, done at a loss.

A decision is the product.

And yet the decision is the one thing the operating stack does not capture, which is why so much of the industry’s margin leaks through decisions that were never written down and therefore never improved.

Consider one shape of this, drawn not from a customer but as an illustration of the mechanism. A contract renewal is approaching. On the surface the account is healthy; every service level is green. But the value actually delivered has never been measured, the relationship runs through a single sponsor, and the last three change requests were quietly scoped below cost. A dashboard shows the green number and moves on. The renewal is signed at the old rate. Eighteen months later the account is underwater and nobody can reconstruct the decision that put it there, because it was never a decision, only a date that arrived. Multiply that by a portfolio and you have the margin squeeze of the previous section, not as a market force but as an accumulation of un-recorded judgements. The same pattern shows up on the way in, in the cost of selling what you cannot deliver, and it is the reason a business can be busy, growing, and quietly unprofitable all at once.

The fragmentation tax

Why does the decision evaporate? Partly because no single system is built to hold it, and a people business runs on a remarkable number of systems. The sell, the hire, the deliver and the pay each have their own software, their own owner, their own definition of the truth, and they do not talk to one another. The reasoning that would have to span them, can we hire for what we are about to sell, at the margin we are promising, has nowhere to live, because there is no system whose job is the whole loop.

The scale of the fragmentation is now well documented. Okta’s annual analysis of its customer base found the average organisation running 101 distinct applications in 2025, the first time the figure has crossed one hundred, with the largest enterprises averaging 231.¹²MuleSoft’s connectivity research puts the number higher still for large enterprises, an average of 897 applications, and finds that only 29% of them are integrated with one another.¹³ The remaining seven in ten are islands. Data is copied by hand, reconciled in spreadsheets, and argued over in meetings whose entire purpose is to decide whose number is right.

The cost of that disconnection is real money, even when it is hard to see. Gartner has long put the average cost of poor data quality, the downstream effect of systems that do not agree, at around USD 12.9m a year per organisation.¹⁴And the deeper cost is not the reconciliation labour but the decisions made on top of numbers nobody fully trusts. KPMG’s work on analytics found that 92% of executives were worried about the reputational risk of using analytics inappropriately, and, more tellingly, that only 35% had a high level of trust in their own organisation’s analytics.¹⁵ When two-thirds of leaders do not trust their own numbers, the decision defaults back to instinct, and the expensive data infrastructure becomes theatre.

For a people business the fragmentation is especially damaging because the important questions are precisely the ones that cross the silos. Whether to take a deal depends on delivery and hiring, which live in other systems. Whether a contract is healthy depends on cost and value, which the service dashboard cannot see. Whether to expand depends on the labour market and the pipeline at once. These are not reporting questions that a better dashboard would answer; they are decisions that require the functions to reason together, which a stack of disconnected tools is structurally unable to do. We have written elsewhere about the hidden cost of disconnected systems; the summary is that the fragmentation is not an IT inconvenience but the reason the most valuable decisions are the ones no system supports.

The AI inflection

Into this arrives artificial intelligence, and it is worth being precise about what it does, because the hype and the fear both obscure it. For people businesses, AI is simultaneously the most significant productivity opportunity in a generation and the most direct threat their model has ever faced, and both are true at once.

The opportunity is measurable. In the most rigorous field study to date, published in the Quarterly Journal of Economics, researchers studied 5,172 customer-support agents given access to a generative-AI assistant and found an average productivity gain of 15%, measured as issues resolved per hour.¹⁶ The distribution matters more than the average: the least-skilled and least-experienced agents improved the most, while the most experienced gained little and occasionally lost a step.¹⁶AI, in other words, compresses the gap between a novice and an expert, which for a labour-intensive business is a profound thing, because it changes what a new hire is worth on day one. Adoption has moved accordingly: McKinsey’s 2025 survey found 72% of organisations using generative AI, up from 33% only a year earlier, with service operations among the leading use cases.¹⁷

The trajectory points further. Gartner forecasts that by 2029 agentic AI, systems that act rather than merely suggest, will autonomously resolve 80% of common customer-service issues without human intervention, cutting operational costs by around 30%.¹⁸ If even directionally right, that reshapes the unit economics of contact-centre outsourcing entirely. But the same analysts counsel against believing the timeline uncritically: Gartner also predicts that more than 40% of agentic-AI projects will be cancelled by the end of 2027, undone by escalating cost, unclear value and inadequate controls.¹⁹ The lesson is not that AI is overhyped; it is that AI deployed without judgement fails expensively, which is a point we will return to.

The threat is structural. The offshore labour-arbitrage model, moving standardised, well-defined, heavily-monitored work to a lower-cost location, was rational precisely because the work was standardised, well-defined and monitorable. Those are exactly the properties that make a task automatable. As the Harvard Business Review put it, AI is rewriting the economics of outsourcing by automating the very characteristics that made arbitrage worthwhile.³⁰A provider whose value proposition is “the same work, cheaper, somewhere else” is selling the thing AI does best. The providers who survive will be the ones who move up the value chain, from doing the work to governing it, from process to outcome, a shift we describe in our guide on BPO versus BTO.

And there is a subtler danger, the one most relevant to this report. An AI that produces a confident answer you cannot trace is not intelligence; it is instinct with a bigger vocabulary. A model that recommends a price, a hire or a staffing level without exposing the evidence behind it asks the organisation to trust it in exactly the way KPMG found executives already do not trust their own analytics.¹⁵ We have argued that AI without evidence creates expensive mistakes, and 2026 is the year that stops being a philosophical objection and starts being an operational and regulatory one. But it is not, in the first place, an AI problem or even a governance problem. It is an architecture problem, and it is worth seeing plainly.

Why today’s software stack cannot solve this

A leader hearing all of this might reasonably ask why the software they already own cannot fix it. They have spent years and budgets on systems, and each one holds something important. Surely, between them, they hold the decision. It is worth walking the stack slowly, because the answer is the argument for everything that follows.

A people business runs on roughly nine systems, and each is good at its job. The CRM holds the deal: its stage, its value, its close date. The applicant tracking system holds the hire. The HR system holds the employment record. The finance system holds the invoice and the cost. The workforce management tool holds the roster. The business intelligence layer holds the dashboards that report on all of them. And around those six sit the three systems no vendor draws on an architecture diagram but every operator actually lives in: Slack, where the reasoning is argued out and then scrolls away; email, where the commitment is buried in a thread; and the spreadsheet, the universal solvent, where the work the other eight cannot do gets done, and then forgotten.

Look at what each system stores, then look for the decision. It is in none of them. The decision to take a deal at a given margin, knowing what delivery would cost, drew on the CRM and the finance system and the workforce tool at once, and was made in a meeting and confirmed in an email. No single system saw all the inputs, and no system kept the output. The reasoning that spanned the stack landed in the one place the stack does not have: a home for the decision itself.

Every system stores a slice of the truth. None of them stores the decision. This is not only a data-quality problem that a better integration would solve, though the integration gap is real enough: the average enterprise runs 897 applications with fewer than a third connected.¹³ It is a category problem. These systems were built to record state, the deal, the hire, the invoice, not to compose a judgement across state, hold the options that were weighed, record who chose and against what evidence, and score the result. You could integrate all nine perfectly and still not have the decision, because none of them was ever designed to hold it. It is why a service dashboard can glow green over an account that is quietly losing money, why the CRM on its own is not enough, and why the hiring system and the finance system never reconcile.

Bolting more tools on top does not close the gap, because they read the same slices. Business intelligence reports what the nine systems already hold, so it inherits their blind spot. And an AI assistant that recommends without exposing its evidence is only a faster version of the meeting whose reasoning scrolled away: confident, unaccountable, and impossible to learn from, which is exactly the failure mode the previous section warned about. The missing layer is not a tenth system alongside the nine. It is a place for the thing that spans them.

That place is what decision intelligence provides, and it is why the discipline has to exist as its own category rather than a feature bolted onto any existing tool. It is also why the two forces still to come, the governance regime and the shift to one operating picture, both point back here. A regulator who asks a people business to justify a decision about a person, how a candidate was screened, why a shift was allocated, is asking for precisely the record none of these nine systems keeps. Governance, next, is less a new burden than a forcing function for something a well-run business needed anyway.

The governance turn

For most of the software era, how a decision was reached was a private matter. If a manager screened a candidate out, or a system flagged an account, or an algorithm set a price, the reasoning stayed inside the building. That era is ending, and people businesses are among the most exposed to its ending, because the decisions they make are decisions about people: who gets hired, who gets the shift, who gets promoted, whose performance is judged.

The European Union’s AI Act is the sharpest expression of the change. It classifies AI systems by risk, and it places squarely in its high-risk category, Annex III, the use of AI in employment, worker management and access to self-employment: recruitment and selection, the filtering and evaluation of candidates, and decisions about promotion, termination, task allocation and the monitoring of performance and behaviour.²⁴ That is not an edge case for a staffing firm or a BPO; it is the core of what they do. And an AI system does not escape the classification because a human signs off at the end, if the system materially shapes the decision, it is in scope. The penalties are calibrated to be felt: up to €35m or 7% of worldwide annual turnover for the most serious, prohibited practices, and up to €15m or 3% for breaches of the high-risk obligations, whichever is higher.²⁵

The timeline has moved, and it is worth being exact, because a lot of commentary is out of date. The Act entered into force in 2024 with a staggered schedule, and the high-risk Annex III obligations were originally due to apply from 2 August 2026. In late 2025 the European Commission proposed a “Digital Omnibus” to simplify and defer parts of the regime; after trilogue, the Council of the EU gave the package its final green light on 29 June 2026.²⁷ The effect is a real extension, not a repeal: standalone high-risk obligations now bind from 2 December 2027, and high-risk systems embedded in regulated products from 2 August 2028.²⁶ Transparency duties for things like chatbots and synthetic media still arrive earlier. The deadline moved; the requirement did not. We keep a running plain-English explainer of what the EU AI Act means for BPOs as the detail settles.

The deadline moved. The requirement did not. By 2027 the question is not whether your AI is clever, but whether you can justify the decision it shaped.

The AI Act does not arrive into a vacuum. GDPR enforcement, which governs the personal data these businesses process by the million, has kept climbing: cumulative fines across Europe reached roughly €6.11bn by March 2026,²⁸ and the volume of reported personal-data breaches rose to 443 a day, the first time above four hundred.²⁹ Recruitment, monitoring and cross-border transfer, precisely the data flows of a people business, feature heavily among the largest cases. The combined message from the regulators is consistent and it points in one direction: a decision that touches a person must increasingly be one you can explain, evidence and defend. An organisation that cannot reconstruct how a decision was made, what data it used, which options it weighed, who chose, is not merely disorganised; by 2027 it is non-compliant. Governance, in other words, is converging on exactly the properties that decision intelligence was built to provide. Our guide on GDPR for BPOs covers the data side in more depth.

What decision intelligence actually is

The forces are now on the table: margins under pressure, decisions that vanish unrecorded, a stack too fragmented to hold a decision, an AI wave that is equal parts opportunity and threat, and a regulatory turn that demands decisions be explainable. They converge on a single requirement. The people business needs to treat its decisions as first-class objects, capture them, evidence them, govern them, and learn from them, rather than let them evaporate into meetings and dashboards. That requirement has a name: decision intelligence.

It is worth saying clearly what it is not, because it is easily confused with two adjacent things. Decision intelligence is not business intelligence: BI reports the past well, in dashboards and metrics, and stops at the edge of the decision. And it is not an AI copilot: a copilot hands you an answer and asks you to trust it. Decision intelligence does the harder thing in between. It composes the available evidence into a recommendation you can interrogate, records the choice a human actually made, and later scores the outcome against it. BI describes; a copilot asserts; decision intelligence decides, and remembers why.

A decision, treated intelligently, has five properties, and it needs all five:

• Evidence carried with its quality. Every fact underneath a decision is labelled by how well it is known, whether it was measured, modelled, inferred, merely stated, or is unmeasured. An assumption is never allowed to impersonate a measurement. This is the evidence hierarchy, and it is what stops a confident guess from being weighed as if it were a fact.
• Options priced against the constraint that binds. Each option carries the earliest date it clears every requirement, not the date someone hopes for. Because the business experiences the constraint that fails, not the average of its constraints, the worst constraint decides. An option that commits to seats the labour market cannot fill is marked blocked, and the system names which constraint blocks it.
• Confidence that is composed, not asserted. Business confidence is computed from the published evidence and weighted by its quality, not typed in by whoever is most senior in the room. The system derives it, and it names the dimension that is dragging it down.
• A recorded human choice. A person always decides. If they overrule the recommendation, the object records who, when, and against what evidence. Overrides are allowed, and remembered, which is how the system learns whether the humans or the model were right.
• A scored outcome. When reality lands, it is compared with what was recommended and what was decided. That score is the institutional memory the organisation currently throws away.

Captured this way, a decision stops being a moment in a meeting and becomes something the business owns. It can be searched, audited, challenged, and, crucially, learned from. This is also, not by coincidence, exactly the shape a regulator will ask for: evidence, options, choice, oversight, record. The discipline that satisfies the auditor is the same one that improves the margin, which is the quiet good news buried in the governance turn. The place where such a decision is made, held and scored, we call a Decision Room; the record it leaves behind is a Decision Object.

From five tools to one operating picture

Decision intelligence is a discipline, not a product, and a discipline needs somewhere to live. For a people business the natural home is the loop its whole economy runs on: sell, hire, deliver, and get paid, each feeding the next. The reason the important decisions evaporate is that this loop is currently spread across systems that cannot see one another. Making decisions intelligent therefore means, in practice, connecting the loop, so that a decision in one function can be reasoned about with the facts from all of them.

The clearest example is the join between the sell and the deliver. A sales team judges a deal on the information in the CRM: stage, value, probability, close date. But in a people business the most important question about a deal lives outside the CRM entirely, can we actually hire and deliver this, at the margin we are quoting, on the date we are promising? A pipeline that does not account for delivery reality is a forecast of intentions, not outcomes. Win rate, the metric most sales teams steer by, is on its own a lagging indicator, it tells you what already happened, not what you can keep, which is why we treat win rate as a lagging indicator rather than a target. Formal proposal processes do not rescue this: even well-run RFP functions win somewhere around 43 to 45% of the bids they choose to enter,³¹ and the ones worth winning are the ones the business can deliver profitably, a judgement the CRM cannot make alone.

The same is true at every other join. Whether an account should be renewed depends on delivered value and true cost, which the service dashboard cannot see. Whether to expand into a new market depends on the pipeline and the labour market at once. Whether a contact-centre programme is healthy depends on metrics, occupancy, shrinkage, attrition, that are routinely read in isolation and mislead when read that way. Occupancy sustained above the mid-eighties, for instance, looks like efficiency and behaves like burnout; the average maximum occupancy target across a large sample of contact centres sits around 83.3% for exactly that reason.²³ None of these are reporting questions. They are decisions that require the functions to reason together, which is the definition of cross-module intelligence.

This is why the answer to fragmentation is not another dashboard bolted on top of the five systems, but a shared spine underneath them, on which each function publishes facts, with their evidence quality, that the others can reason with. The goal is not one giant application replacing five; it is one operating picture the five can share, so that the decision, the thing that was always spread across them, finally has a place to live. That is the architecture behind the ONX platform, and the philosophy behind it is set out in full in Enterprise Decision Intelligence.

A decision-intelligence maturity model

Organisations do not adopt decision intelligence in a single step, and it is useful to have a map of the journey. The model below is our own framework, offered as a way to locate where a people business sits today and what the next move looks like. It is a synthesis of the patterns we see, not a measured benchmark; treat it as a lens, not a league table.

Most people businesses today sit between stage 1 and stage 2. They have invested heavily in reporting, each function can produce a competent dashboard, and some have begun to connect their systems. Very few have reached the point where a decision is captured as a governed object, and fewer still have closed the loop so that outcomes teach the next decision. That gap, between reporting the past and governing the future, is where the margin discussed at the start of this report is being lost, and where the opportunity now sits.

The reason the top of the model matters is that it compounds. A single well-governed decision is useful. A thousand of them, each scored against its outcome, is an asset no competitor can buy, because it is built from your own history. If commitments with a certain profile consistently deliver below forecast, the next recommendation with that profile says so. This is the loop that turns a pile of decisions into enterprise intelligence, and it is why the destination is not “better dashboards” but an organisation that gets measurably better at deciding, quarter after quarter.

What to do in 2026

The forces in this report are not going to relent. Margins will stay contested, AI will keep advancing on both fronts, and the governance requirements will bind in 2027 whether or not an organisation is ready. Waiting is itself a decision, and, on the evidence here, an expensive one. There are five moves that a people business can make now, and none of them require a moonshot.

• Instrument your worst constraint, not your average. The business fails at the constraint that binds, hiring you cannot do, delivery you cannot staff, so build the habit of asking, of every commitment, which single thing has to be true for this to work, and whether it is. Read decisions against the binding constraint, not the comfortable average.
• Connect the sell to the deliver first. Of all the joins in the loop, this is the one that leaks the most margin. Before a deal is committed, it should be read against what can actually be hired, delivered and afforded. This is where connecting even two systems pays for itself.
• Label the quality of your evidence. Start distinguishing, explicitly, between what you have measured and what you have assumed. Most bad decisions are not made on bad data; they are made on assumptions wearing the costume of data. The evidence hierarchy is a discipline you can begin applying in a spreadsheet on Monday.
• Adopt AI where evidence is attached, and be sceptical where it is not. The productivity gains are real, especially for your least-experienced people. But a recommendation you cannot trace is a liability, operationally and, from 2027, legally. Insist that any AI in a consequential decision exposes the evidence beneath it.
• Start recording decisions as objects now, ahead of the deadline. The governance regime arriving in 2027 asks for exactly what good decision making needs anyway: evidence, options, oversight, a record. Building that habit now is not a compliance cost; it is the thing that also recovers the margin. Do it early and the regulation becomes a formality rather than a scramble.

None of these is a technology project in the first instance; each is a change in how a decision is treated. The technology, a shared spine on which functions publish evidence and decisions are made, governed and scored, is what makes the habit sustainable at scale. But the habit comes first, and the habit is available to any organisation willing to stop throwing its reasoning away.

Methodology & a note on honesty

A word on how this report was made, because the subject demands it. A document about the importance of traceable, evidence-labelled decisions has no business making untraceable, unevidenced claims of its own.

Every quantitative figure here is drawn from a named, dated, public source, analyst firms, regulators, peer-reviewed research and public-company filings, and each is listed in the References below. Where sources disagree, as they do on the size of the outsourcing market and on contact-centre attrition, we have said so and given the range rather than picking the most dramatic number. Where a widely-repeated statistic could not be traced to a primary source, we have left it out; the frequently-quoted figure of “USD 10,000 to 20,000 to replace a contact-centre agent,” for example, is attributed across the industry to McKinsey and to SQM Group but is very hard to pin to an original publication, so we have not used it as a load-bearing claim, only noted the caution.

A report about traceable decisions has no business making untraceable claims of its own.

One element is explicitly our own: the maturity model in the section above is a framework, a synthesis of observed patterns, not a measured benchmark, and it is labelled as such. The single renewal example in “The decision that never gets stored” is likewise illustrative, offered to show the shape of a mechanism, not to report a specific customer outcome. We hold ourselves to the same evidence hierarchy we recommend to others: measured facts are cited, modelled or illustrative content is named as such, and assumptions are not dressed up as measurements.

This is the first annual edition. We intend to publish The State of Decision Intelligence each year, tracking how the discipline, and the forces pushing people businesses toward it, actually move. Treat the 2026 edition as the baseline: the value compounds when there is a 2027 to read it against.

This report will be revised as the numbers move, particularly the AI-adoption figures and the EU AI Act timeline, both of which are changing quickly. 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 status by being corrected, not by being defended.

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. Business Process Outsourcing (BPO) Market Size & Outlook, 2030 · Grand View Research, 2025. Global BPO market USD 302.6bn (2024), projected USD 525.23bn by 2030 at a 9.8% CAGR.
2. Global Staffing Market Estimates and Forecasts · Staffing Industry Analysts (SIA), 2025. Global staffing revenue ~USD 620bn (2024); the US ~USD 184bn and the UK ~USD 50.8bn.
3. Consulting Growth to Double After Sluggish 2024, Research Finds (reporting Source Global Research) · Bloomberg Tax, 2025. Global consulting grew just 2.9% in 2024; Source forecasts ~6% growth toward ~USD 110bn across its tracked markets in 2025.
4. Managed Services Market Size, Share & Growth Report, 2032 · Fortune Business Insights, 2025. Managed services USD 297.2bn (2024), projected USD 878.7bn by 2032 at a 15.0% CAGR.
5. Fourth Quarter and Fiscal Year 2025 Results · Concentrix Corporation, 2026. Non-GAAP operating margin 12.8% (FY2025), down from 13.7% (FY2024); a Q4 goodwill impairment drove GAAP operating margin negative.
6. 2024 Annual Results · Teleperformance, 2025. Recurring EBITA margin 15.0% (FY2024); 2025 guidance implied broadly flat margins.
7. Teleperformance shares fall 13% after Concentrix results · MarketScreener, 2025. A sector-wide margin scare; Teleperformance shares fell roughly 13%.
8. 2024 Global Outsourcing Survey · Deloitte, 2024. Cost reduction remains the primary outsourcing driver for more than 70% of organisations; buyers increasingly evaluate fully-loaded cost.
9. Decision making in the age of urgency · McKinsey & Company, 2019. Only 20% of organisations say they excel at decision making; 61% say most decision-making time is used ineffectively; inefficiency costs a typical Fortune 500 ~530,000 manager-days (~USD 250m) a year.
10. Three keys to faster, better decisions · McKinsey & Company, 2019. Respondents who called their decisions fast were 1.98x more likely to also call them high quality; “winning” organisations were about twice as likely to report top-quartile financial returns from their decisions.
11. Measuring decision effectiveness · Bain & Company, 2020. A 95%+ correlation between decision effectiveness and financial performance; top-quintile decision firms averaged ~6 percentage points higher total shareholder return.
12. Businesses at Work 2025 · Okta, 2025. The average organisation now runs 101 apps (the first time above 100), and the largest organisations average 231.
13. Connectivity Benchmark Report 2025 · Salesforce MuleSoft, 2025. The average enterprise uses 897 applications, of which only 29% are integrated.
14. How to Improve Your Data Quality · Gartner, 2021. Poor data quality costs organisations an average of USD 12.9m a year (Gartner, attributed to Melody Chien).
15. Guardians of Trust (Forrester Consulting study for KPMG International) · KPMG, 2018. 92% of executives worry about the reputational damage of inappropriate analytics use; only 35% report high trust in their own organisation’s analytics.
16. Generative AI at Work, The Quarterly Journal of Economics 140(2), 889-942 · Brynjolfsson, Li & Raymond, 2025. A study of 5,172 customer-support agents found a 15% average productivity gain, concentrated among the least-experienced; the most experienced gained little.
17. The State of AI · McKinsey & Company, 2025. 72% of organisations report using generative AI, up from 33% a year earlier; 88% use AI in at least one function.
18. Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029 · Gartner, 2025. Forecast issued March 2025, with a projected 30% reduction in operational costs.
19. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 · Gartner, 2025. Cancellations driven by escalating costs, unclear business value and inadequate risk controls.
20. What Metrigy’s Latest AI Data Reveals About Contact Center Staffing · Metrigy, 2024. Contact-centre agent turnover rose from 21.8% (2022) to 28.1% (2023) to 31.2% (2024).
21. Global Contact Center Survey · Deloitte Digital, 2023. Reported average annual agent attrition around 52%; AI-centric contact centres reported being 85% more profitable.
22. Professional Services Maturity Benchmark 2025 · SPI Research, 2025. Average billable utilisation fell to 66.4% (2025) from 68.9% (2024), the lowest in the survey’s history; high-performing firms reach ~81.2%.
23. What Is the Ideal Occupancy Rate for a Contact Centre? · Call Centre Helper, 2023. An analysis of more than 160,000 calculations put the average maximum occupancy target at 83.3%.
24. Annex III: High-Risk AI Systems (Regulation (EU) 2024/1689) · EU Artificial Intelligence Act, 2024. Employment, worker management and access to self-employment (including recruitment, screening, evaluation and performance monitoring) are classified as high-risk.
25. Article 99: Penalties (Regulation (EU) 2024/1689) · EU Artificial Intelligence Act, 2024. Up to €35m or 7% of worldwide annual turnover for prohibited practices; up to €15m or 3% for most other breaches, whichever is higher.
26. AI Act Update: EU Resolves to Change Rules and Extend Deadlines · Latham & Watkins, 2026. The Digital Omnibus defers Annex III high-risk obligations to 2 December 2027 (standalone systems) and 2 August 2028 (embedded systems).
27. Artificial intelligence: Council gives final green light to simplify and streamline rules · Council of the EU, 2026. Final adoption of the AI Act simplification package, 29 June 2026.
28. GDPR Enforcement Tracker Report: Numbers and Figures · CMS, 2026. Cumulative GDPR fines reached roughly €6.11bn by March 2026.
29. Personal data breaches in Europe reach 443 per day in dramatic 22% jump · DLA Piper, 2026. Breach notifications rose to 443 a day across Europe, the first time above 400.
30. AI Is Rewriting the Economics of Outsourcing · Harvard Business Review, 2026. Generative AI automates precisely the standardised, monitorable tasks that made offshore labour arbitrage rational.
31. RFP Response Trends and Benchmarks · Loopio, 2025. Average RFP win rates cluster around 43 to 45%; general B2B win rates are far lower and vary widely by deal size.