AI agents for business: an operating model for scaling without headcount

Most large surveys now agree on the shape of the problem: companies have tried agents, but few have scaled them. McKinsey's state of AI survey for 2025 found that 62% of respondents say their organisations are at least experimenting with AI agents. Twenty-three per cent report that they are scaling an agentic AI system, and a further 39% have begun experimenting. In any given business function, no more than 10% of respondents say their organisations are scaling agents. The distance between trying and scaling is wide. The finding is consistent with what many leadership teams report privately: a number of promising demonstrations and very few processes that have been redesigned around them.

The wider picture is similar. McKinsey reports that 88% of respondents use AI regularly in at least one function, up from 78% a year earlier, but only 39% attribute any level of EBIT impact to AI. In the UK, the Office for National Statistics reports that business use of AI rose from around 12% in late 2023 to around 35% by June 2026, while only 10% of businesses with 10 or more employees say they use it extensively. Adoption is broad and shallow, which is the classic profile of a technology still waiting for its operating model. Gartner's January 2025 poll of 3,412 webinar attendees tells a similar story about investment: only 19% had made significant investments in agentic AI.

For a chief executive, the implication is that licences and pilots are not evidence of progress. A company can show a dozen agent experiments and still have no measurable change in cost, speed or revenue. Boards should be wary of dashboards that count tools deployed rather than output delivered. The meaningful question is how many workflows now run materially faster or cheaper, who owns each one, and what the next quarter's target is. Few firms can answer it today.

Start with a baseline. List every AI and agent use in the business, the person who owns it, the metric it is meant to move and the latest reading. Stop anything with no owner or metric within thirty days. What remains is your real portfolio, and it will usually be smaller than expected. That portfolio, not the pilot count, is the starting point for deciding where to scale and where to stop. Publish the register to the leadership team so that everyone sees the same picture, and so that duplicated experiments can be merged.

Evidence from the companies furthest ahead supports a workflow-first approach. Boston Consulting Group's 2025 research, based on a global survey of 1,250 senior executives and AI decision makers, found that agents already account for 17% of total AI value in 2025, and that this share is expected to reach 29% by 2028. Future-built companies, which BCG says are 5% of companies globally, allocate 15% of their AI budgets to agents. Agents are a growing slice of value, but still a minority, and the leaders spend selectively. The BCG survey covered nine industries and more than 25 sectors, which gives the figures reasonable breadth across the economy.

That pattern matters for the mid-market because it contradicts the instinct to launch something grand. Even the most advanced firms do not put most of their AI budget into agents; they reserve it for areas where autonomy pays. A growing company with a smaller budget and a thinner technical bench has even more reason to be selective. Each deployment should be small enough to explain on one page, specific enough to measure, and valuable enough that a manager will defend it when things go wrong. For the same reason, we would not recommend a separate budget line labelled agents; fund each workflow from the function that will benefit.

The implication is that the unit of planning is the workflow, not the function or the technology. A sales team is not a candidate for an agent; the step that researches a prospect, drafts a first message and logs the result is. A service desk is not a candidate; the triage of routine queries and the drafting of standard replies are. Describing work at that level exposes whether it has clear inputs, rules and outputs, which is the real test of whether an agent can do it. Writing the workflow as a sequence of steps also reveals which of them involve judgement that should stay with a person.

Choose three workflows and define for each the unit of output, such as qualified leads researched per week, queries resolved without escalation, or invoices matched without error. Record the current cost and time per unit. Set a target for the first ninety days. Agree who can switch the agent off. If you cannot define the unit of output, the workflow is not ready for an agent, and the preparatory work on process is itself valuable.

The market is crowded with claims, and buyers need a filter. Gartner estimates that only about 130 of the thousands of agentic AI vendors are real, a reminder that relabelled automation is common. At the same time, Gartner predicts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024. Agents will arrive inside your software whether or not you plan for them. The practical consequence is that vendor selection should follow workflow selection, and never the other way round.

The selection logic is practical. Volume matters because an agent's fixed set-up cost needs many repetitions to pay back. Rule clarity matters because the less ambiguous the instructions, the more reliable the output. Cost of error matters because a drafting mistake in an internal summary is trivial, while a wrong figure in a customer invoice is not. Data availability completes the picture: an agent cannot reason over records it cannot reach. Workflows that score well on all four are the sensible first candidates. A fifth test, which is easy to overlook, is whether the output has an obvious owner who will notice when it is wrong.

Applied to a growing company, the first wave usually sits in repetitive, rules-bound tasks: research and enrichment of prospects, drafting and logging of follow-ups, first-line responses to routine service queries, document intake, reconciliation and onboarding checklists. High-volume work with ambiguous rules, such as complaint handling, is better supported than automated. Low-volume work with clear rules may need only a template or a script. The exhibit on this page positions typical workflows on those two axes, and the position should drive the sequence. Treat the quadrant as a sequencing tool rather than a verdict, since workflows move towards the upper right as rules are written down and data is cleaned.

Score ten candidate workflows from one to five on each of the four criteria and rank them. Run the top three in shadow mode first, so the agent drafts and a person approves, and compare its output with the human baseline. Move to supervised autonomy only when quality is stable. Avoid buying a platform until you know which workflows it must support, since vendor claims are far easier to test against a specific job than against a general promise.

Governance is where many programmes quietly fail. A Deloitte survey of 3,235 information technology and business leaders from 24 countries found that only 21% say their organisations have a mature governance model in place for agentic AI, even though 74% expect their companies to be using AI agents at least moderately by 2027. Ambition is outrunning control. For a company of any size, the risk is not only error but an inability to say what an agent did, why it did it and who approved it. A smaller company faces the same exposure on a smaller scale, with fewer people to notice a fault.

Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027. Its January 2025 poll of 3,412 webinar attendees shows how tentative most organisations still are: 19% had made significant investments in agentic AI, 42% conservative investments and 8% none, with the remaining 31% taking a wait-and-see approach or unsure. In our assessment, cancellation is most likely where there is no named owner, no measure of success and no route for escalation. Those are management failures rather than technical ones. Wait-and-see is a rational posture where the use case is unclear, but it becomes costly when competitors have a measured head start.

The implication is that an AI operating model should look like a small version of how the company already manages people. Every agent needs an owner who is accountable for its output, a scorecard with quality and volume measures, a probation period in which a human checks its work, defined limits on what it may decide or spend, and an escalation route to a person. It also needs a record of the actions it takes so that errors can be traced and corrected quickly. Where a vendor supplies the agent, require the same register entry and audit trail as for one you build yourself.

Create an agent register with one line per agent: purpose, owner, data it can access, actions it may take, approval limits, scorecard and review date. Review it monthly in the same meeting that reviews hiring and performance. Add a simple rule that no agent may take an irreversible action, such as paying a supplier or sending a contract, without human approval until it has met its quality threshold for a full quarter.

Leaders are divided on what agents mean for jobs, and the evidence is mixed. In McKinsey's survey, 32% of respondents expect their workforce to decrease over the next year, 43% expect no change and 13% expect an increase. Microsoft's 2025 Work Trend Index found that a third of leaders (33%) are considering headcount reductions, while many are also considering hiring for AI-specific roles. Hiring is not disappearing; it is changing shape. Equally, the same survey shows that leaders are planning for new roles, such as people who train, supervise and secure agents.

Microsoft also reports that 81% of leaders expect agents to be moderately or extensively integrated into their company's AI strategy, and that 46% say their companies are already using agents to fully automate workflows or processes. These are self-reported figures from leaders, and they should be read with caution: enthusiasm in a survey is not the same as a deployed, measured system. Still, they show the direction of travel and the level of competitive expectation that growing companies face. Treat the Microsoft numbers as an indicator of leadership intent rather than a measure of operational maturity.

For a mid-market chief executive, the practical lesson is to avoid promising a headcount outcome in advance. A company that is growing can often absorb the work released by agents into new volume, which means hiring is avoided rather than roles removed. That is a better story internally and a more accurate one. It also removes the incentive for managers to hide results. Capacity released is the measure to track, with a separate decision about how to use it. Be explicit with staff about what the programme is for; ambiguity breeds resistance and quietly degrades the data that agents rely on.

Define capacity released in hours per week for each deployed agent, validated by the process owner and reviewed monthly. Decide in advance how the released capacity will be used: handling growth without new hires, shifting people to higher-value work, or reducing cost. Set a rule that open vacancies in an affected team are reviewed against measured capacity before they are advertised. This creates a disciplined link between agent output and hiring decisions without turning the programme into a redundancy exercise.

Size and sector shape adoption, and the UK data show it clearly. The Office for National Statistics reported in June 2026 that 28% of businesses with 0 to 9 employees were using AI, compared with 49% of those with 250 or more. Use ranged from 58% in information and communication to 13% in construction. Smaller firms are not shut out, but the largest are ahead, and sector culture matters as much as size. A growing company should benchmark itself against its own sector and the next size band up. Both statistics describe businesses of different sizes and should be read as ranges, not targets for any individual company.

McKinsey's figures on scaling point the same way. In its 2025 survey, nearly half of respondents from companies with more than $5 billion in revenue had reached the scaling phase, compared with 29 percent of those with less than $100 million in revenue. Larger firms have more data, more specialists and budgets for governance. Mid-market companies have a different advantage: shorter decision chains and fewer legacy systems. The advantage is real only if processes are documented and data is accessible, which is not always the case. The gap between those two groups is a reminder that scale confers resources, but not necessarily speed.

The implication is that preparation, not purchasing, is the constraint. An agent that works from inconsistent customer records, undocumented rules or scattered files will produce inconsistent results, and the company will blame the technology. Cleaning up the underlying process often delivers a gain of its own. Where the process is already tidy, the same company can run a first agent in weeks, and a board that sees a measured result in one quarter is far more likely to support the second. A tidy workflow with an accountable owner will usually beat a sophisticated tool pointed at an untidy process.

Before selecting a tool, document the target workflow end to end: inputs, steps, rules, exceptions, systems touched and the person who currently decides. Fix obvious data gaps, such as duplicate contacts or missing fields. Then run a four- to six-week pilot with a named owner and a baseline. Treat the first agent as a test of the company's readiness as much as of the technology, and write down what you learned for the next one.

For companies operating in or selling into Saudi Arabia and the United Arab Emirates, the benchmark is moving quickly. HP's Work Relationship Index 2026, reported by Middle East AI News, found that 60% of Saudi workers use AI agents, against 57% in the UAE. These are worker surveys rather than company deployments, but they indicate how normal agent use has become in daily work in the region, and how fast employees there expect their employers to provide capable tools. Treat the regional picture as context for competitive positioning, not as proof that any individual company is behind.

KPMG's Saudi Arabia Tech Report 2026 points to similar momentum among organisations. It found that 99% of Saudi respondents report active investment in agentic AI, and that 46% have already deployed AI in production at scale across multiple use cases. The survey covered 2,500 technology leaders globally, of whom only 70 were in Saudi Arabia, so the national figures should be treated as indicative rather than definitive. Even so, they suggest a market in which intent is high and execution is being funded. Where a global survey reports Gulf results from small samples, we would rather say so than overstate precision.

The implication for UK and European mid-market companies is competitive as well as commercial. Customers, partners and rivals in the Gulf may operate with faster response times and lower service costs than expected, and local procurement teams may ask what AI capability a supplier brings. Companies entering the region will meet counterparts who are comfortable with agent-assisted work, and will be judged on responsiveness. Meanwhile, Deloitte's finding that 74% of leaders expect to use agents at least moderately by 2027 indicates that the same expectations are spreading elsewhere. Responsiveness is one of the first things a new counterpart notices, and agent-assisted triage is one way to improve it.

Use these figures to calibrate ambition, not to chase them. If your sector or customers sit in the Gulf, ask what response times and service levels counterparts now expect, and test whether a targeted agent can help you match them. Prioritise customer-facing workflows with clear rules, such as enquiry triage and scheduling. Keep governance in step with speed, since regional regulators and enterprise buyers are likely to ask how automated decisions are overseen.

Karter McKinley supports AI and technology implementation by selecting the first workflows, defining the baseline and unit of output, and setting up the agent register, scorecards and governance routines. We work with existing systems and data where possible and avoid new subscriptions unless a workflow justifies them. The aim is measured output from each deployment, with a clear decision to scale, change or stop.