AI for Business — Strategy, Adoption, and ROI

The 2026 honest baseline

In May 2026, three numbers define the AI-in-business landscape:

• ~80% of organizations use generative AI in at least one function (McKinsey Global AI Survey, March 2026)
• ~60% report no enterprise-wide financial impact
• ~5% see "substantial ROI" (IBM)

Adoption is universal. Value capture is rare. The gap is what this article is about.

Why most AI projects underperform

Three failure modes recur across postmortems:

1. No baseline measurement

Teams ship AI features without recording the pre-AI baseline. Six months later, no one can prove the AI changed anything. Without baselines you can't answer "did this work?"

2. AI sitting outside the system of record

A copilot that lives in a separate app, not wired into the CRM / ticketing / project management system where outcomes are tracked. Adoption looks high in usage logs; outcomes are invisible.

3. Treated as an experiment, not a function

No named owner, no KPIs, no governance, no budget authority. The project drifts until quiet death.

The shippers (12% of agent pilots that reach production, per DigitalApplied 2026 data) share three traits:

• Named owner with budget authority — 94% of shippers
• Automated evals on every change — 87%
• Human-in-the-loop checkpoints — 74%

The BCG 10-20-70 rule

The most-cited and most accurate adoption framework comes from BCG: spend your effort budget as

• 10% on algorithms (model choice, prompt engineering, fine-tuning)
• 20% on data and technology (RAG, integrations, infrastructure)
• 70% on people, process, and change management (training, redesigned workflows, role redefinition, governance)

Engineering teams instinctively invert this — they spend 70% on algorithms and 10% on change management. That's why their projects underperform. The technology is commoditized; the people work isn't.

Depth over breadth

BCG's 2025 research found leaders prioritize 3.5 use cases on average; laggards prioritize 6.1. The leaders' yield was roughly 2× the ROI.

The reason: each use case requires its own change-management overhead — workflow redesign, training, governance, KPIs, evaluation. Splitting 70% of effort across six projects means each gets 12%. Across three, each gets 23%. Below ~20% of effort, change management fails.

Practical translation: pick three to four use cases per year, go deep, finish them, measure impact, then expand.

What's actually returning money (2026)

Payback time by function, from the DigitalApplied 2026 enterprise data:

By industry, banking and insurance lead at 47% production adoption. Healthcare and government lag below 20% — regulation, integration complexity, and risk tolerance explain most of the gap.

The early-wins pattern is stable: coding copilots, support deflection, sales prospecting. Start there if you don't have an obvious project champion elsewhere.

Buy vs build vs partner

A pragmatic decision tree:

• BUY horizontal capabilities. Chat (Claude/ChatGPT), coding (Cursor/Claude Code), meetings (Granola/Fathom), search (Perplexity), automation (Zapier). The foundation-model layer moves too fast to out-build; you'll waste resources catching up.

• PARTNER for domain-specific workflows where vendors don't have your proprietary data. Healthcare ambient scribing, legal document analysis, vertical CRM features. Vendor + your data + your workflows.

• BUILD only for differentiated workflows where proprietary data is the moat. Don't build a chatbot. Don't build a coding agent. Build the thing that competitors can't replicate because they don't have your data.

Governance you can actually use

The minimal governance stack:

• AI registry — every AI feature/agent in production, with owner, purpose, last-eval date
• Eval suites — automated test sets for each AI feature; run on every change
• Human-in-the-loop policies for destructive actions (writes to production data, customer-facing communications, irreversible decisions)
• Privacy + data-handling review — what data goes to which model provider, with what retention
• Incident response — who's paged when a model misbehaves; documented mitigation playbooks

Below this is exposure. Above this is bureaucracy. Find the line.

The CEO is the AI decision-maker

A 2026 BCG number worth internalizing: ~75% of CEOs are the AI decision-maker at their companies. AI is no longer "an IT topic." Corporate AI spend is projected to roughly double in 2026 — from ~0.8% of revenue to ~1.7%.

What this means practically: if your AI champion is a director-level person without CEO sponsorship, your initiative is at risk of being defunded the next time priorities shift. Get the CEO involved early — they're the only person who can sustain investment across function silos.

Where to start

For a company beginning AI adoption in 2026:

1. Pick three use cases with clear baseline metrics and named owners.
2. Train your top quartile — the 25% of employees who'll champion adoption with their peers.
3. Wire AI into the systems of record — CRM, ticketing, project management — not standalone apps.
4. Set up evals + monitoring before launch.
5. Plan a 12-month measurement window — 90 days isn't enough for most use cases to prove out.
6. Budget 70% of effort for change management — train, redesign workflows, document.

Companies that do this end up in the ~5% with substantial ROI. Companies that don't end up in the ~60% with no measurable impact.