Executive Summary
Nearly every organisation now uses AI somewhere in the business, yet only a small minority can point to money on the bottom line. That gap — between using AI and profiting from it — is the strategic problem of 2026. It matters because spending has doubled, CEOs have taken personal ownership of the results, and the performance gap between leaders and laggards is widening rather than closing. This article is written for executives, business owners and department leaders who need to decide where AI investment goes next. The biggest takeaway: stop buying tools and start redesigning the work around them.
Estimated reading time: 12 minutes
Key Takeaways
- Broad AI use has become table stakes. Measurable financial return has not, and that difference is now the competitive dividing line.
- Companies seeing real returns concentrate on a handful of use cases rather than spreading budget across dozens of pilots.
- Most AI money should go to people and process change, not to models. The proven split is roughly 10% algorithms, 20% data and technology, 70% people and process.
- AI agents are moving into production, but most pilots still fail — and governance, not capability, is usually the blocker.
- The EU AI Act’s transparency rules are already enforceable and reach any company whose AI outputs touch EU residents.
- Track financial outcomes such as cost per transaction and cycle time. Usage statistics tell you nothing about value.
The Strategic Challenge: Everyone Has AI, Few Have Returns
Ask a room of executives whether their organisation uses AI and almost every hand goes up. Ask which of them can name the profit impact and most hands come down.
That is the honest state of enterprise AI in 2026. According to McKinsey’s State of AI 2026 research, 88% of organisations use AI in at least one business function and 72% use generative AI specifically — up from 33% two years earlier. It is the fastest adoption ramp of any enterprise technology on record.
The returns have not followed. Research from BCG and KPMG puts the share of enterprises reporting measurable, at-scale financial returns at just 5–8%. McKinsey’s own survey finds only 39% of organisations reporting any EBIT impact at all. In PwC’s 29th Global CEO Survey, 56% of CEOs said AI delivered neither revenue growth nor cost reduction in the previous twelve months.
This is what the research calls the adoption-to-value gap, and it has three practical consequences for leaders.
It changes what “being behind” means. Two years ago, the risk was not using AI. Now that almost everyone does, using it is no longer a differentiator. The risk is spending at market rate and getting nothing measurable back.
It puts the strategy itself under scrutiny. In a 2026 survey by WRITER, 75% of executives admitted their AI strategy is “more for show” than a real internal guide. A strategy nobody uses is a document, not a plan.
It raises the cost of drift. Corporate AI spending has doubled from roughly 0.8% to 1.7% of revenue, according to BCG’s AI Radar 2026, and 79% of enterprises report cost overruns. Doing AI badly is now expensive enough to notice.
Why AI Changes the Situation
Earlier technology waves rewarded procurement. Buy the system, install it, train people on it, and value followed reasonably predictably.
AI does not behave that way. The tool is generic; the value comes from the work you rebuild around it. That is why two companies can license the same software and one sees a 3x return while the other sees a slightly faster version of the process it already had.
Future Signal Tip If you can describe your AI initiative without mentioning a single process that changed, you have bought software, not built a strategy.
Who This Strategy Is For
This approach applies most directly to:
- Enterprise organisations already spending meaningful money on AI without a clear return picture.
- Growing companies (roughly 50–500 people) at the point where informal tool adoption is becoming an unmanaged cost.
- Executive teams who now own AI outcomes personally rather than delegating them to IT.
- Department leaders in customer support, finance, legal and sales, where the documented returns are strongest.
- Any business whose AI outputs reach EU residents, regardless of where it is headquartered.
When This Strategy Is Not Yet Appropriate
Be honest about readiness. This approach is premature if:
- Your data is scattered, inconsistent or undocumented. Nearly half of enterprises name data quality as their top deployment obstacle, and no amount of model quality fixes a broken input.
- Nobody owns the outcome. Without a named owner, AI work drifts back into side-project status.
- The organisation is mid-restructure. Workflow redesign requires stable processes to redesign.
- You are a very small team where one or two well-chosen tools genuinely are the whole strategy. Concentrate on using them well and revisit this when headcount and process complexity grow.
The Current Landscape
Three shifts define the business environment going into the second half of 2026.
AI has become a board-level accountability, not a technology programme. BCG’s AI Radar 2026, based on 2,400 executives including 640 CEOs across 16 markets, found that nearly three-quarters of CEOs are now their organisation’s main AI decision-maker — double the share of a year earlier. Half believe their job depends on AI paying off. More than 90% intend to keep investing at current or higher levels even if it does not pay off within the year.
That combination — personal accountability plus committed spending — means the pressure to show results is going to intensify, not ease.
Agents are moving from demo to deployment. The share of enterprises with at least one AI agent in production has climbed from 9% in 2024 to 31% in 2026, and 80% of enterprise applications shipped in the first quarter of 2026 embedded at least one agent.
| Agent adoption signal | 2024 | 2025 | 2026 |
|---|---|---|---|
| Enterprise apps embedding at least one agent | 33% | 58% | 80% |
| Enterprises with an agent in production | 9% | 19% | 31% |
| Deployments coordinating three or more agents | 1% | 6% | 22% |
| Enterprises with a named agent owner | 11% | 27% | 56% |
Source: Digital Applied, compiling Gartner, S&P Global and McKinsey data.
Regulation has become operational. The EU AI Act’s transparency obligations under Article 50 took effect on 2 August 2026, covering chatbot disclosure, AI content marking and deepfake labelling. Obligations for standalone high-risk systems were pushed to 2 December 2027 under the Digital Omnibus amendments approved in June 2026, with product-embedded high-risk systems following in August 2028.
The delay is a scheduling change, not a reprieve. The transparency rules are live now, and maximum fines reach 7% of global annual turnover or €35 million — above GDPR’s 4% ceiling.
Strategic Opportunities
1. Concentrate Investment Instead of Spreading It
The single clearest pattern in the 2026 research is that focus beats breadth. BCG finds leading companies work an average of 3.5 use cases against 6.1 for everyone else, put more than 80% of AI investment into reshaping core functions rather than peripheral experiments, and expect 2.1x greater return than peers.
The mechanism is straightforward. Every additional pilot splits attention, change-management capacity and executive patience. Three initiatives that reach production beat ten that stall in evaluation.
2. Deploy Agents Where the Payback Is Already Documented
Some use cases have enough deployment history to be treated as known quantities rather than experiments.
| Use case | Enterprises deploying | Median ROI |
|---|---|---|
| Customer support automation | 62% | 3.4x |
| Software engineering assistance | 58% | 2.9x |
| Document and contract analysis | 47% | 3.1x |
| Knowledge management and internal search | 41% | 2.6x |
| Data analysis and BI augmentation | 39% | 2.4x |
| Compliance and regulatory monitoring | 27% | 2.7x |
| Sales enablement and content generation | 44% | 2.2x |
Source: Presenc.ai, Enterprise AI Adoption Statistics 2026. Figures are aggregated across surveyed enterprises and will vary by organisation.
Median payback across agent deployments sits at 5.1 months, and median time from pilot to production has fallen to 4.2 months from eleven months in 2024. Those numbers give you a reasonable benchmark: if a pilot has not reached production in six months, something structural is wrong.
3. Redesign Workflows Rather Than Adding Tools
This is where most of the value sits, and where most organisations have not gone. Only around 21% of organisations have redesigned workflows for generative AI. McKinsey reports that top-performing companies see roughly $3 back for every $1 invested, with core profit up about 20% on average over three to five years.
The distinction matters more than it sounds:
TOOL ADDITION
Existing process ──► same steps ──► one step slightly faster
│
└─► marginal, hard-to-measure gain
WORKFLOW REDESIGN
Existing process ──► remove steps ──► reassign human judgement
└─► rebuild handoffs ──► measure new baseline
│
└─► compounding, measurable gain4. Apply the 10-20-70 Split
BCG’s allocation model — 10% to algorithms and models, 20% to data and technology infrastructure, 70% to people, processes and cultural change — is the pattern behind the strongest performers. Companies following it report 1.7x revenue growth, 2.7x return on invested capital and 1.6x EBIT margin against laggards.
Most budgets get this backwards, loading spend onto licences and infrastructure while treating training and process work as an afterthought.
5. Use Tiered Models to Control Cost
Routing routine work to lower-cost models and reserving premium models for high-stakes decisions can cut infrastructure costs by 40–60%. With median monthly model spend growing more than sevenfold year over year, this stops being a technical optimisation and becomes a margin question.
Strategic Risks
The Pilot Graveyard
Roughly 88% of agent pilots never reach production. The blockers reported by leaders are evaluation gaps (64%), governance friction (57%) and model reliability (51%) — two of the three are organisational, not technical.
More uncomfortably, 54% of C-suite executives told WRITER’s 2026 survey that AI adoption is “tearing their company apart,” and only 29% of companies report significant returns despite 59% investing at least $1 million annually.
Governance That Has Not Kept Pace
Only one in five companies has a mature governance model for autonomous agents. Globally, just 12.4% have a human oversight policy, and 48% of those have not documented the operational processes that would make it meaningful. Gartner’s assessment is that more than 40% of agent projects will be abandoned by 2027 if governance and ROI fundamentals go unaddressed.
⚠️ Callout: Governance is a scaling prerequisite, not paperwork The organisations struggling most with agents are rarely the ones with weaker technology. They are the ones that cannot answer basic questions: who approved this agent, what is it allowed to do without a human, and where is the record of what it did? Answer those three questions before you scale, not after an incident forces you to.
Regulatory Exposure
The EU AI Act reaches beyond Europe. If your AI outputs affect EU residents, you are in scope regardless of where you are headquartered. Compliance costs for large enterprises are estimated at $8–15 million, with third-party certification adding $50,000 or more per system.
The more immediate pressure is commercial. EU-based customers are writing AI Act provisions into RFPs, due diligence questionnaires and contracts, which means supply chain compliance arrives through your sales pipeline before any regulator contacts you.
Note that enforcement practice is still new — no significant enforcement actions had been reported as of August 2026 — so the practical severity of penalties remains projected rather than observed.
The Skills Gap
Insufficient worker skills is the biggest single barrier to integrating AI into existing workflows, according to Deloitte. Only 53% of organisations are educating the broader workforce, 48% have upskilling strategies, and fewer than one in three companies has upskilled even a quarter of its people.
Data Quality, Security and Lock-In
Around 48% of enterprises name data problems as their top deployment challenge. On security, 2026’s threat landscape increasingly pits automated attacks against automated defences, which raises the cost of manual-only security operations. And concentrating on a single vendor’s agent framework creates switching costs worth understanding before you standardise.
Decision Box: What Should You Do Right Now?
Act now if: you already have AI in production somewhere, you can name a process worth rebuilding, and you have an executive willing to own the outcome.
Pilot first if: you have identified two or three high-value use cases but have no baseline metrics yet. Establish those before deploying anything.
Fix foundations first if: your data is inconsistent, no one owns AI outcomes, or you have no inventory of the AI systems already in use.
Monitor if: you are a small team where current tools meet current needs. Revisit when process complexity or headcount grows.
Regardless of the above: if your AI outputs reach EU residents, Article 50 transparency compliance is not optional and is already in force.
Implementation Roadmap
The research supports a twelve-month sequence with go/no-go gates rather than a continuous rollout.
Stage 1 — Assess Readiness (Months 1–3)
Inventory every AI system in use, in development and in procurement. More than half of organisations lack this basic starting point, which makes both governance and cost control impossible. Risk-classify each system against the EU AI Act tiers, and capture baseline metrics for the processes you intend to change: processing time, error rate, cost per transaction, hours spent.
The baseline is the part most teams skip and later regret, because without it you cannot prove improvement.
Stage 2 — Define Objectives and Ownership
Select three to five use cases with clear financial metrics attached. Name an AI agent owner or operations lead — 56% of enterprises that scale successfully have done this. Set the target: not “improve efficiency” but “reduce cost per support ticket by 30% within nine months.”
Stage 3 — Select Technology
Choose tools against the workflow you have designed, not the other way around. Plan for a tiered model approach from the start. Confirm what audit logging and access control the platform supports, because retrofitting governance is considerably harder than requiring it upfront.
Stage 4 — Pilot (Months 3–6)
Deploy minimum viable agents with 100% human review of outputs. Build observability early: tracing, cost monitoring and quality evaluation. Test failure deliberately with malformed inputs, data outages and prompt injection attempts.
Then apply a genuine gate. Proceed only if accuracy meets your threshold and you have identified and mitigated at least three failure modes. Cancelling at this gate is a success, not a failure.
Stage 5 — Measure in Limited Production (Months 6–9)
Move to a controlled 20–30% of real workload. Monitor daily, wire audit logging into your security monitoring, isolate agent execution, and document the ROI evidence you will need to justify the next stage.
Stage 6 — Scale (Months 9–12)
Expand to additional departments, introduce tiered model routing for cost, build orchestration where multiple agents genuinely need to coordinate, and begin reporting enterprise-level financial KPIs rather than project-level ones.
Common Mistakes
Measuring adoption instead of value. Licence counts and weekly active users are easy to collect and tell you nothing. They persist because they make progress reports look good. Replace them with cost, cycle time and error rate against your baseline.
Running too many pilots. This usually comes from wanting to keep every department happy. The result is that no initiative gets enough change-management support to reach production. Fund fewer things properly.
Treating AI as an IT project. The 70% of value that sits in people and process cannot be delivered by a technology function alone. Business process owners need to be accountable, not merely consulted.
Scaling before governance exists. Teams defer governance because it feels like drag during a pilot. It becomes drag at exactly the moment you want to move fastest. Define autonomy tiers against business risk — not agent capability — before production.
Skipping the baseline. Without a before, there is no after. This is the most common reason a genuinely successful project cannot survive budget review.
Assuming the regulation does not apply. Extraterritorial reach catches a lot of companies who assumed a non-EU headquarters exempted them.
Success Metrics
Split your measurement into three layers. Operational metrics tell you whether the system works; financial metrics tell you whether it matters; governance metrics tell you whether it is safe to scale.
| Layer | Metric | Reference point |
|---|---|---|
| Operational | Cycle time reduction | 25%+ on automated workflows |
| Operational | Agent accuracy | ≥70% in pilot, ≥85% in production |
| Operational | User adoption | ≥50% of target workforce |
| Operational | Pilot to production | ≤4.2 months (2026 median) |
| Financial | Payback period | ≤5.1 months (2026 median for agents) |
| Financial | Return on AI investment | ≥2.4x median; 5.1x+ is top quartile |
| Financial | EBIT impact | Measurable within 12–24 months |
| Governance | AI systems inventoried and classified | 100% |
| Governance | Audit trail coverage of agent actions | 100% |
| Governance | Workforce AI fluency training | ≥50% of workforce |
Benchmarks are 2026 market medians drawn from Presenc.ai, BCG and Forrester data. Treat them as reference points for your own targets, not guarantees.
Readiness Checklist
Work through this before committing budget to a scaled deployment:
- Every AI system in the business is inventoried and risk-classified
- Three to five priority use cases are selected, with financial metrics attached
- Baseline measurements are captured for each target process
- A named owner is accountable for AI outcomes at executive level
- Autonomy tiers are defined against business risk, with human approval required for high-stakes actions
- Audit logging captures agent actions and feeds security monitoring
- Article 50 transparency measures are in place where AI outputs reach EU residents
- A workforce AI fluency programme is running, not merely planned
- Budget allocation approximates 10% models, 20% data and technology, 70% people and process
Future Outlook: The Next 12–24 Months
What is confirmed. The EU AI Act’s high-risk obligations arrive on 2 December 2027, with product-embedded systems following in August 2028. That is a fixed date to plan against.
What the industry expects. Deloitte finds more than 60% of organisations expect to deploy AI agents within two years, up from 23% today. Gartner projects 40% of enterprise applications will embed task-specific agents by the end of 2026. Agent spending is tracking toward roughly $1.4 trillion by 2027 on IDC and McKinsey estimates, and physical AI use is projected to rise from 58% to 80% within two years. In some industries, AI budgets are on track to overtake cloud infrastructure as the largest IT line item by 2027.
What is projection rather than evidence. IMD faculty anticipate 10–20% compression in traditional middle-management roles by the end of 2026 and the emergence of AI-native departments where 40–60% of activity runs autonomously. These are informed projections, not verified findings. MIT Sloan’s Davenport and Bean separately expect the AI investment bubble to begin deflating in 2026 — a qualitative judgement rather than a modelled forecast.
Our reading. These last two predictions point in opposite directions, and both could be partly right. A correction in speculative AI spending would not contradict continued growth in deployments with documented payback. If anything, tighter capital makes the discipline described in this article more valuable, not less.
The Future Signal
The interesting shift in 2026 is not technical. It is that AI has stopped being a technology decision and become an operating-model decision — and most organisations are not structured to make those.
Technology decisions have owners, budgets and procurement processes. Operating-model decisions require someone with authority to change how work is divided between people and systems, and to accept that some roles and processes will not survive the change. That authority usually sits higher than where AI programmes currently report.
This is why the gap between leaders and laggards is widening rather than closing. The leaders are not buying better models — the models are broadly available to everyone. They are making harder organisational decisions faster.
What deserves attention: the workflows in your business where a person mostly moves information between systems, applies a routine rule, or waits for someone else’s output. That is where value concentrates.
What can be safely ignored: most model release announcements, benchmark leaderboards, and the recurring debate about whether AI is overhyped. Both the enthusiasts and the sceptics are arguing about the technology. The returns are being decided by organisational design.
The companies that will look well-positioned in 2028 are the ones treating this as a management problem now.
What Businesses Should Do Next
In priority order:
- Run an AI inventory this quarter. You cannot govern, cost or improve what you have not catalogued. This is the cheapest high-value action available.
- Name an owner with real authority. Not a coordinator — someone accountable for the financial outcome.
- Cut your use case list to three to five. Choose from the categories with documented returns and clear internal sponsors. Stop or park the rest.
- Capture baselines before deploying anything else. Processing time, error rate, cost per transaction, hours spent.
- Address Article 50 transparency now if any AI output reaches EU residents. Chatbot disclosure and content marking are enforceable today.
- Rebalance the budget toward people and process. If more than a third of your AI spend is going to licences and infrastructure, the allocation is working against you.
- Define autonomy tiers and audit logging before scaling agents, not after.
- Start workforce fluency training for the teams whose work will change first.
Frequently Asked Questions
How much should we be spending on AI? Market average is around 1.7% of revenue, but the average is a poor target. Concentration matters more than volume — the research consistently shows companies spending less with tighter focus outperforming companies spending more across many pilots.
Should we build our own AI systems or buy? For the use cases with documented returns — support, document analysis, coding assistance — buying is usually faster and cheaper. Building makes sense where the workflow is genuinely proprietary and central to how you compete. In either case, the 70% of effort in people and process is unavoidable.
We ran a pilot and it did not work. Should we try again? Look at why it stopped first. The dominant blockers are evaluation gaps and governance friction rather than model quality, and both are fixable. If you never established a baseline, you may not actually know whether it worked.
How do we know if a use case is worth pursuing? Three tests: the process is high-volume and rule-heavy, you can measure it today, and a business owner outside IT wants it changed. If any one is missing, deprioritise it.
Does the EU AI Act apply to us if we are outside the EU? If your AI outputs affect EU residents, yes. Many non-EU companies also encounter it commercially, through EU customers embedding AI Act requirements in contracts and RFPs.
What do we do about employees worried AI will replace them? Be specific rather than reassuring. Vague promises erode trust faster than honest scope. IBM’s research found employees across age groups support increased AI use at two to three times the rate they oppose it — the anxiety is usually about being excluded from the decision, not about the technology.
Is agentic AI ready for production use? In some areas, clearly yes — 31% of enterprises have agents in production with a 5.1-month median payback. But 88% of pilots still fail, and some credible commentators think broad agent readiness is still a few years away. Treat it as ready in proven categories with human oversight, and experimental elsewhere.
How long before we see returns? Current benchmarks put pilot to production at around four months and agent payback at around five. EBIT-level impact typically takes twelve to twenty-four months. If someone promises enterprise-wide transformation in a quarter, adjust expectations.
Conclusion
The strategic challenge of 2026 is not whether to use AI. That question has been answered by nearly everyone. The challenge is that broad use has produced narrow returns, and the organisations closing that gap are doing it through organisational discipline rather than better technology.
The recommended approach is consistent across the evidence: concentrate on a few use cases, redesign the work rather than adding tools to it, put most of the budget into people and process, govern agents before scaling them, and measure financial outcomes instead of activity.
The biggest takeaway is that AI value is created by the decisions around the tool, not by the tool. That is genuinely harder than procurement — but it is also why the advantage is durable once you build it.
The next logical step is the smallest one: inventory what you already have, pick the three processes worth rebuilding, and measure them before you change anything. Most organisations discover they are already spending enough. They are simply spreading it too thin to see.
