Every big company is "doing AI". Very few are getting paid for it.
Three numbers tell the story.
- 78% of organizations used AI in at least one business function in 2024, up from 55% a year earlier. (McKinsey, The state of AI, 2024)
- Just 1% of leaders describe their company as "mature" on AI deployment. (McKinsey)
- 74% of companies struggle to achieve and scale value from AI investment. (BCG, 2024)
You can buy ChatGPT Enterprise for 10,000 seats. That doesn't move your P&L. What moves the P&L is workflow redesign — which is the work almost nobody is actually doing.
This guide is about how we think about that work at Kiri Media. We're an AI-native agency; the ops pattern we ship for our own clients transfers directly to in-house teams at bigger companies. The punchline: the gains are real, the number is in the 20–30% range on the right tasks, and the path there is almost entirely about workflow design and change management — not tool licensing.
The productivity numbers are unambiguous.
The best-run controlled studies have converged on similar ranges:
- 40% better performance on in-scope tasks when consultants use GPT-4 — and 12.2% more tasks completed, 25.1% faster on average. (Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School / BCG, 2023)
- 55.8% faster task completion for developers using GitHub Copilot on a coding task. (Peng et al., MIT/Stanford, 2023)
- 14% overall productivity gain in customer support — 35% for new hires — when agents worked alongside a generative AI assistant. (Brynjolfsson, Li, Raymond, NBER Working Paper 31161)
- 0.1–0.6 percentage points added to annual labor productivity for the next decade — an economic impact of $2.6–4.4 trillion per year. (McKinsey, The economic potential of generative AI, 2023)
So the ceiling is real. What's harder is getting there.
Three failure modes we see everywhere.
Tool deployment is not workflow change
You roll out Copilot or Gemini to the organization, measure adoption via license seats, and declare victory. A year later, almost every employee uses it for the same thing: drafting emails. That's a 5% individual gain, maybe. It doesn't compound into a P&L number.
The "jagged frontier" is invisible
The Harvard/BCG study made this concept legible: AI is brilliant at some tasks, actively harmful on others, and the boundary is not intuitive. Consultants who used AI on tasks outside its frontier performed worse than the control group. Most training programs teach the tool. Your teams learn where the button is — not where the edge is.
No agent infrastructure
Enterprise-grade AI productivity is not one person asking ChatGPT a question. It's an agent that runs on a schedule, pulls data from four systems, drafts the output, flags edge cases for human review, logs its work, and retries on failure. That requires engineering. Buying seats doesn't get you that; it gets you a chat window.
The workflows that deliver 20–30%.
The agent workflows where we consistently see team-level gains share the same structure:
- One narrow end-to-end job. Not "help with writing" — but "draft the weekly product-launch brief from our calendar, pull pricing from Shopify, enrich with competitor data, and land in Notion for review." The agent owns the whole loop, not a step.
- Humans at the decision points, not the production steps. Humans approve. They don't type the draft.
- Observability from day one. Every agent run is logged. Every output is reviewable. You see what the agent did, why, and what it cost.
- Iterated weekly, not quarterly. The agent's prompt, tool list, and review rules are living documents — owned by the team, not the vendor.
This is the shape we build at Kiri Media — for paid-media audits, SEO agent workflows, creative review, post-campaign reporting. The same pattern works inside a 500-person marketing org.
Four layers of enterprise AI that compound.
Think of it as a stack. Each layer multiplies the one below.
Chat, autocomplete, inline suggestions
Expected gain: 10–15% on individual tasks. This is where roughly 95% of enterprises have stopped. It's useful, but it caps out at the level of the individual — it doesn't change how work moves through the organization.
One job, end-to-end
Expected gain: 25–40% on the task. The agent drafts the contract, the brief, the report. A human reviews and approves. The hand-off cost disappears for that task.
Multi-step, multi-system, multi-approval
Expected gain: compounds over Layer 2. The entire SEO audit → content brief → publish cycle. Inbound lead → enrichment → qualified routing. Time savings compound because hand-offs disappear across the whole chain, not just one step.
Monitoring, anomaly detection, triage, first-pass analysis
Expected gain: frees 20–30% of human capacity permanently. Running 24/7, escalating only when it matters. Most enterprises don't build this — and it's where the actual productivity unlock lives.
Cost per layer is roughly flat. Gain is roughly exponential. Most enterprises build Layer 1 and stop; the ones pulling ahead are building Layers 3 and 4.
What this looks like concretely.
A recent Nordic e-commerce client — YMYL vertical, 2,000+ SKUs, editorial team of four, dev team of six, roughly 12,000 monthly organic visits. Before we started:
- Product meta descriptions were written manually, when someone had time. Forty percent were empty.
- Editorial could publish two articles per week, on a good week. Author attribution was missing on four of every six pieces — a compliance problem in their vertical.
- Technical SEO issues (schema, canonicals, hreflang) piled up because no one had time to hunt them down.
An agent owns the long-tail production now. It drafts meta descriptions against the brand's voice guidelines. It flags articles missing author attribution. It opens pull requests for trivial technical fixes. The team reviews and ships. Editorial capacity effectively 1.3×'d — without a hire.
The agent cost runs around 0.5% of the freed headcount cost. That's not a typo.
How to actually get there.
Enterprises that will win this cycle aren't buying more AI tools. They are:
- Picking three high-leverage workflows — not the easiest, but the ones with the highest hand-off cost today.
- Redesigning each as an end-to-end agent loop with explicit approval gates.
- Building the observability layer before the agent ships, not after.
- Training teams on the frontier — where AI is strong, where it's dangerous — not on the tool.
- Measuring in output per week, not in license adoption.
Steps 2 and 3 are where most organizations lack the engineering muscle. That's the gap we help close at Kiri Media. We've spent three years shipping agent workflows for growth, SEO, and creative at client scale — the same architecture transfers directly into an enterprise ops function.
If this is the problem you're solving.
Send a note with a few sentences about your company, your teams, and the workflows you'd most want to agent-ify. First call is a scoping conversation — no deck, no pitch.